enigma-memory 0.1.13 → 0.1.15

package/scripts/run-standard-memory-benchmarks.mjs

@@ -1,1070 +1,1070 @@
-#!/usr/bin/env node
-import { createReadStream } from 'node:fs';
-import { mkdir, readFile, writeFile } from 'node:fs/promises';
-import { basename, dirname, resolve } from 'node:path';
-import { fileURLToPath } from 'node:url';
-import { createHash } from 'node:crypto';
-import { performance } from 'node:perf_hooks';
-import { StringDecoder } from 'node:string_decoder';
-import { estimateTextTokens } from '../packages/optimizer/src/index.js';
-
-export const STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA = 'enigma.standard_memory_benchmark_suite.v1';
-
-export const STANDARD_MEMORY_BENCHMARK_METHODS = Object.freeze([
-  Object.freeze({
-    id: 'full_context',
-    label: 'Full context',
-    boundary: 'Supplies every parsed memory record for each query; no provider API or model answer generation.',
-    uses_top_k: false,
-  }),
-  Object.freeze({
-    id: 'recency_last_n',
-    label: 'Recency last N',
-    boundary: 'Supplies the most recent local memory records up to --top-k.',
-    uses_top_k: true,
-  }),
-  Object.freeze({
-    id: 'keyword_filter',
-    label: 'Keyword filter',
-    boundary: 'Supplies local memory records whose public-safe deterministic tokens overlap the query, capped by --top-k.',
-    uses_top_k: true,
-  }),
-  Object.freeze({
-    id: 'enigma_relevance',
-    label: 'Enigma relevance',
-    boundary: 'Uses deterministic query-aware relevance features over local public-safe memory metadata and content tokens, then ranks locally without provider APIs.',
-    uses_top_k: true,
-  }),
-]);
-
-const LOCOMO_SOURCE_URL = 'https://raw.githubusercontent.com/snap-research/locomo/main/data/locomo10.json';
-const LONGMEMEVAL_SOURCE_URLS = Object.freeze([
-  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_oracle.json',
-  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json',
-  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_m_cleaned.json',
-]);
-
-const QUERY_RELEVANCE_STOPWORDS = new Set([
-  'about',
-  'after',
-  'again',
-  'against',
-  'also',
-  'and',
-  'any',
-  'are',
-  'assistant',
-  'because',
-  'been',
-  'before',
-  'being',
-  'between',
-  'can',
-  'could',
-  'current',
-  'does',
-  'from',
-  'has',
-  'have',
-  'how',
-  'into',
-  'its',
-  'latest',
-  'more',
-  'most',
-  'number',
-  'own',
-  'owns',
-  'please',
-  'should',
-  'that',
-  'the',
-  'their',
-  'then',
-  'there',
-  'these',
-  'they',
-  'this',
-  'use',
-  'using',
-  'was',
-  'what',
-  'when',
-  'where',
-  'which',
-  'who',
-  'whose',
-  'why',
-  'with',
-  'would',
-]);
-
-function parseArgs(argv = process.argv.slice(2)) {
-  const options = { top_k: 5 };
-  for (let index = 0; index < argv.length; index += 1) {
-    const arg = argv[index];
-    if (arg === '--locomo') {
-      options.locomo = requiredFlagValue(argv, index, arg);
-      index += 1;
-    } else if (arg === '--longmemeval') {
-      options.longmemeval = requiredFlagValue(argv, index, arg);
-      index += 1;
-    } else if (arg === '--max-locomo-qa') {
-      options.max_locomo_qa = positiveInteger(requiredFlagValue(argv, index, arg), arg);
-      index += 1;
-    } else if (arg === '--max-longmemeval-items') {
-      options.max_longmemeval_items = positiveInteger(requiredFlagValue(argv, index, arg), arg);
-      index += 1;
-    } else if (arg === '--top-k') {
-      options.top_k = positiveInteger(requiredFlagValue(argv, index, arg), arg);
-      index += 1;
-    } else if (arg === '--out') {
-      options.out = requiredFlagValue(argv, index, arg);
-      index += 1;
-    } else if (arg === '--help' || arg === '-h') {
-      options.help = true;
-    } else {
-      throw new Error(`Unknown option ${arg}`);
-    }
-  }
-  return options;
-}
-
-function requiredFlagValue(argv, index, flag) {
-  const value = argv[index + 1];
-  if (value === undefined || value.startsWith('--')) throw new Error(`${flag} requires a value`);
-  return value;
-}
-
-function positiveInteger(value, name) {
-  const number = Number(value);
-  if (!Number.isInteger(number) || number <= 0) throw new Error(`${name} must be a positive integer`);
-  return number;
-}
-
-function optionalPositiveInteger(value, name) {
-  if (value === undefined || value === null) return undefined;
-  return positiveInteger(value, name);
-}
-
-function publicFileName(path) {
-  return path === undefined || path === null ? undefined : basename(String(path));
-}
-
-async function readJsonWithSha256(path) {
-  const raw = await readFile(path, 'utf8');
-  return {
-    data: JSON.parse(raw),
-    sha256: createHash('sha256').update(raw).digest('hex'),
-  };
-}
-
-function isJsonWhitespace(char) {
-  const code = char.charCodeAt(0);
-  return code === 0x20 || code === 0x0a || code === 0x0d || code === 0x09 || code === 0xfeff;
-}
-
-function createTopLevelArraySampler(maxItems, name) {
-  const items = [];
-  let started = false;
-  let closed = false;
-  let collecting = false;
-  let doneCollecting = false;
-  let expectSeparator = false;
-  let depth = 0;
-  let inString = false;
-  let escaped = false;
-  let current = '';
-
-  function fail(message) {
-    throw new Error(`${name} sample-mode JSON parse failed: ${message}`);
-  }
-
-  function finishItem() {
-    items.push(current);
-    current = '';
-    collecting = false;
-    expectSeparator = true;
-    if (items.length >= maxItems) doneCollecting = true;
-  }
-
-  return {
-    write(text) {
-      for (const char of text) {
-        if (!started) {
-          if (isJsonWhitespace(char)) continue;
-          if (char !== '[') fail('expected a top-level array');
-          started = true;
-          continue;
-        }
-
-        if (doneCollecting) continue;
-
-        if (closed) {
-          if (!isJsonWhitespace(char)) fail('found trailing data after the top-level array');
-          continue;
-        }
-
-        if (!collecting) {
-          if (isJsonWhitespace(char)) continue;
-          if (expectSeparator) {
-            if (char === ',') {
-              expectSeparator = false;
-              continue;
-            }
-            if (char === ']') {
-              closed = true;
-              continue;
-            }
-            fail('expected a comma or closing bracket between items');
-          }
-          if (char === ']') {
-            closed = true;
-            continue;
-          }
-          if (char !== '{' && char !== '[') fail('expected each sampled item to be an object or array');
-          collecting = true;
-          current = char;
-          depth = 1;
-          continue;
-        }
-
-        current += char;
-        if (inString) {
-          if (escaped) {
-            escaped = false;
-          } else if (char === '\\') {
-            escaped = true;
-          } else if (char === '"') {
-            inString = false;
-          }
-          continue;
-        }
-        if (char === '"') {
-          inString = true;
-        } else if (char === '{' || char === '[') {
-          depth += 1;
-        } else if (char === '}' || char === ']') {
-          depth -= 1;
-          if (depth < 0) fail('encountered an unmatched closing bracket');
-          if (depth === 0) finishItem();
-        }
-      }
-    },
-    get done() {
-      return doneCollecting;
-    },
-    finish() {
-      if (!started) fail('empty input');
-      if (!doneCollecting && collecting) fail('ended inside a sampled item');
-      if (!doneCollecting && inString) fail('ended inside a string');
-      if (!doneCollecting && !closed) fail('ended before the top-level array closed');
-      return JSON.parse(`[${items.join(',')}]`);
-    },
-  };
-}
-
-async function readJsonArraySampleWithSha256(path, maxItems, name) {
-  const hash = createHash('sha256');
-  const decoder = new StringDecoder('utf8');
-  const sampler = createTopLevelArraySampler(maxItems, name);
-  for await (const chunk of createReadStream(path)) {
-    hash.update(chunk);
-    if (!sampler.done) sampler.write(decoder.write(chunk));
-  }
-  if (!sampler.done) {
-    const tail = decoder.end();
-    if (tail.length > 0) sampler.write(tail);
-  }
-  return {
-    data: sampler.finish(),
-    sha256: hash.digest('hex'),
-  };
-}
-
-async function readLongMemEvalJsonWithSha256(path, maxItems) {
-  if (maxItems === undefined) return readJsonWithSha256(path);
-  try {
-    return await readJsonArraySampleWithSha256(path, maxItems, 'LongMemEval');
-  } catch (error) {
-    if (error instanceof Error && error.message === 'LongMemEval sample-mode JSON parse failed: expected a top-level array') {
-      return readJsonWithSha256(path);
-    }
-    throw error;
-  }
-}
-
-function normalizeDatasetArray(data, name) {
-  if (Array.isArray(data)) return data;
-  if (data && typeof data === 'object') {
-    for (const key of ['data', 'items', 'examples', 'samples']) {
-      if (Array.isArray(data[key])) return data[key];
-    }
-  }
-  throw new TypeError(`${name} dataset must be a JSON array or object containing an array`);
-}
-
-function addMeaningfulToken(tokens, token) {
-  if (token.length < 3) return;
-  if (!/[a-z]/u.test(token)) return;
-  if (QUERY_RELEVANCE_STOPWORDS.has(token)) return;
-  tokens.add(token);
-}
-
-function stemToken(token) {
-  if (token.length > 5 && token.endsWith('ing')) {
-    let stem = token.slice(0, -3);
-    if (stem.length > 3 && stem.at(-1) === stem.at(-2)) stem = stem.slice(0, -1);
-    return stem;
-  }
-  if (token.length > 4 && token.endsWith('ed')) {
-    let stem = token.slice(0, -2);
-    if (stem.length > 3 && stem.at(-1) === stem.at(-2)) stem = stem.slice(0, -1);
-    return stem;
-  }
-  if (token.length > 4 && token.endsWith('ies')) return `${token.slice(0, -3)}y`;
-  if (token.length > 4 && token.endsWith('es')) return token.slice(0, -2);
-  if (token.length > 3 && token.endsWith('s') && !token.endsWith('ss')) return token.slice(0, -1);
-  return token;
-}
-
-function addStemmedMeaningfulToken(tokens, token) {
-  addMeaningfulToken(tokens, token);
-  const stem = stemToken(token);
-  addMeaningfulToken(tokens, stem);
-  if (token.endsWith('ed') || token.endsWith('ing')) addMeaningfulToken(tokens, `${stem}e`);
-}
-
-function meaningfulTokensFrom(value) {
-  const tokens = new Set();
-  if (value === undefined || value === null) return tokens;
-  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
-    const token = match[0];
-    addMeaningfulToken(tokens, token);
-    if (token.includes('-') || token.includes('_')) {
-      for (const part of token.split(/[-_]+/u)) addMeaningfulToken(tokens, part);
-    }
-  }
-  return tokens;
-}
-
-function stemmedMeaningfulTokensFrom(value) {
-  const tokens = new Set();
-  if (value === undefined || value === null) return tokens;
-  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
-    const token = match[0];
-    addStemmedMeaningfulToken(tokens, token);
-    if (token.includes('-') || token.includes('_')) {
-      for (const part of token.split(/[-_]+/u)) addStemmedMeaningfulToken(tokens, part);
-    }
-  }
-  return tokens;
-}
-
-function stemmedTokenSequenceFrom(value) {
-  const sequence = [];
-  if (value === undefined || value === null) return sequence;
-  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
-    const token = match[0];
-    const parts = token.includes('-') || token.includes('_') ? token.split(/[-_]+/u) : [token];
-    for (const part of parts) {
-      const stem = stemToken(part);
-      if (stem.length >= 3 && /[a-z]/u.test(stem) && !QUERY_RELEVANCE_STOPWORDS.has(stem)) sequence.push(stem);
-    }
-  }
-  return sequence;
-}
-
-function tokenOverlapScore(queryTokens, record) {
-  if (queryTokens.size === 0) return 0;
-  const recordTokens = meaningfulTokensFrom(record.content);
-  let score = 0;
-  for (const token of queryTokens) if (recordTokens.has(token)) score += 1;
-  return score;
-}
-
-function normalizedTagSessionToken(value) {
-  return String(value ?? '').toLowerCase().replace(/^session[-_:]?/u, '').replace(/[^a-z0-9]+/gu, '');
-}
-
-function roleHintsFrom(value) {
-  const hints = new Set();
-  const text = String(value ?? '').toLowerCase();
-  for (const match of text.matchAll(/\b(?:assistant|user|system|human|agent|speaker[-_\s]?[a-z0-9]+)\b/gu)) {
-    const compact = match[0].replace(/\s+/gu, '_');
-    hints.add(compact);
-    for (const token of stemmedMeaningfulTokensFrom(compact)) hints.add(token);
-  }
-  return hints;
-}
-
-function sessionHintsFrom(value) {
-  const hints = new Set();
-  const text = String(value ?? '').toLowerCase();
-  for (const match of text.matchAll(/\b(?:session|sess)\s*[-_:]?\s*([a-z0-9]+)\b/gu)) hints.add(match[1]);
-  for (const match of text.matchAll(/\bd\s*[-_:]?\s*(\d+)\b/gu)) hints.add(`d${Number(match[1])}`);
-  for (const match of text.matchAll(/\bsession[-_]([a-z0-9]+)\b/gu)) hints.add(match[1]);
-  return hints;
-}
-
-function dateHintsFrom(value) {
-  const hints = new Set();
-  const text = String(value ?? '').toLowerCase();
-  for (const match of text.matchAll(/\b(?:19|20)\d{2}\b/gu)) hints.add(match[0]);
-  for (const match of text.matchAll(/\b\d{4}[-/]\d{1,2}(?:[-/]\d{1,2})?\b/gu)) hints.add(match[0].replace(/\D+/gu, '-'));
-  for (const match of text.matchAll(/\b(?:jan(?:uary)?|feb(?:ruary)?|mar(?:ch)?|apr(?:il)?|may|jun(?:e)?|jul(?:y)?|aug(?:ust)?|sep(?:t(?:ember)?)?|oct(?:ober)?|nov(?:ember)?|dec(?:ember)?)\b/gu)) hints.add(match[0].slice(0, 3));
-  for (const match of text.matchAll(/\b(?:today|yesterday|tomorrow|recent|recently|latest|newest|current|previous|last|earliest|oldest)\b/gu)) hints.add(match[0]);
-  return hints;
-}
-
-function recordSessionTokens(record) {
-  const tokens = new Set();
-  const values = [record.session_id, record.dialog_id, record.turn_id, ...(record.tags ?? [])];
-  for (const value of values) {
-    if (value === undefined || value === null) continue;
-    const normalized = normalizedTagSessionToken(value);
-    if (normalized) tokens.add(normalized);
-    for (const hint of sessionHintsFrom(value)) tokens.add(normalizedTagSessionToken(hint));
-  }
-  return tokens;
-}
-
-function countSetIntersection(left, right) {
-  let count = 0;
-  for (const value of left) if (right.has(value)) count += 1;
-  return count;
-}
-
-function phraseAndProximityScore(querySequence, recordSequence) {
-  if (querySequence.length < 2 || recordSequence.length < 2) return 0;
-  const recordBigrams = new Set();
-  const positions = new Map();
-  for (let index = 0; index < recordSequence.length; index += 1) {
-    const token = recordSequence[index];
-    if (!positions.has(token)) positions.set(token, []);
-    positions.get(token).push(index);
-    if (index > 0) recordBigrams.add(`${recordSequence[index - 1]}\u0000${token}`);
-  }
-  let score = 0;
-  for (let index = 1; index < querySequence.length; index += 1) {
-    const previous = querySequence[index - 1];
-    const current = querySequence[index];
-    if (previous === current) continue;
-    if (recordBigrams.has(`${previous}\u0000${current}`)) {
-      score += 10;
-      continue;
-    }
-    const leftPositions = positions.get(previous);
-    const rightPositions = positions.get(current);
-    if (!leftPositions || !rightPositions) continue;
-    let near = false;
-    for (const left of leftPositions) {
-      for (const right of rightPositions) {
-        if (Math.abs(left - right) <= 6) {
-          near = true;
-          break;
-        }
-      }
-      if (near) break;
-    }
-    if (near) score += 4;
-  }
-  return score;
-}
-
-function hasRecencyIntent(hints) {
-  for (const hint of hints) {
-    if (hint === 'recent' || hint === 'recently' || hint === 'latest' || hint === 'newest' || hint === 'current' || hint === 'previous' || hint === 'last') return true;
-  }
-  return false;
-}
-
-function enigmaRelevanceScore(query, record) {
-  const question = String(query.question ?? '');
-  const queryTokens = stemmedMeaningfulTokensFrom(question);
-  const categoryTokens = stemmedMeaningfulTokensFrom(`${query.category ?? ''} ${query.question_type ?? ''}`);
-  const contentTokens = stemmedMeaningfulTokensFrom(record.content);
-  const metadataTokens = stemmedMeaningfulTokensFrom(`${record.kind ?? ''} ${(record.tags ?? []).join(' ')}`);
-  for (const token of stemmedMeaningfulTokensFrom(`${record.role ?? ''} ${record.session_id ?? ''} ${record.dialog_id ?? ''} ${record.turn_id ?? ''}`)) {
-    metadataTokens.add(token);
-  }
-
-  const contentMatches = countSetIntersection(queryTokens, contentTokens);
-  const metadataMatches = countSetIntersection(queryTokens, metadataTokens);
-  const categoryMatches = countSetIntersection(categoryTokens, metadataTokens) + countSetIntersection(categoryTokens, contentTokens);
-  const roleMatches = countSetIntersection(roleHintsFrom(question), roleHintsFrom(record.role));
-  const querySessionHints = sessionHintsFrom(question);
-  const recordSessions = recordSessionTokens(record);
-  const sessionMatches = countSetIntersection(querySessionHints, recordSessions);
-  const queryDateHints = dateHintsFrom(question);
-  const temporalMatches = countSetIntersection(queryDateHints, dateHintsFrom(`${record.content} ${(record.tags ?? []).join(' ')}`));
-  const temporalRecencyScore = hasRecencyIntent(queryDateHints) ? Math.min(6, Math.log2(record.ordinal + 2)) : 0;
-  const phraseScore = phraseAndProximityScore(stemmedTokenSequenceFrom(question), stemmedTokenSequenceFrom(record.content));
-
-  return {
-    record,
-    score: (contentMatches * 12)
-      + (metadataMatches * 4)
-      + (categoryMatches * 3)
-      + (roleMatches * 18)
-      + (sessionMatches * 16)
-      + (temporalMatches * 8)
-      + temporalRecencyScore
-      + phraseScore,
-    contentMatches,
-    metadataMatches,
-    categoryMatches,
-    roleMatches,
-    sessionMatches,
-    temporalMatches,
-    temporalRecencyScore,
-    phraseScore,
-  };
-}
-
-function compareRecordId(left, right) {
-  return String(left.id).localeCompare(String(right.id));
-}
-
-function compareRecencyDesc(left, right) {
-  if (left.ordinal !== right.ordinal) return right.ordinal - left.ordinal;
-  return compareRecordId(left, right);
-}
-
-function rankedByOverlap(records, query) {
-  const queryTokens = meaningfulTokensFrom(query);
-  if (queryTokens.size === 0) return [];
-  const scored = [];
-  for (const record of records) {
-    const score = tokenOverlapScore(queryTokens, record);
-    if (score > 0) scored.push({ record, score });
-  }
-  scored.sort((left, right) => {
-    if (left.score !== right.score) return right.score - left.score;
-    return compareRecencyDesc(left.record, right.record);
-  });
-  return scored.map((item) => item.record);
-}
-
-function rankedByEnigmaRelevance(records, query) {
-  const scored = [];
-  for (const record of records) {
-    const item = enigmaRelevanceScore(query, record);
-    if (item.score > 0) scored.push(item);
-  }
-  if (scored.length === 0) return [...records].sort(compareRecencyDesc);
-  scored.sort((left, right) => {
-    if (left.score !== right.score) return right.score - left.score;
-    if (left.phraseScore !== right.phraseScore) return right.phraseScore - left.phraseScore;
-    if (left.contentMatches !== right.contentMatches) return right.contentMatches - left.contentMatches;
-    if (left.roleMatches !== right.roleMatches) return right.roleMatches - left.roleMatches;
-    if (left.sessionMatches !== right.sessionMatches) return right.sessionMatches - left.sessionMatches;
-    if (left.temporalMatches !== right.temporalMatches) return right.temporalMatches - left.temporalMatches;
-    if (left.temporalRecencyScore !== right.temporalRecencyScore) return right.temporalRecencyScore - left.temporalRecencyScore;
-    if (left.metadataMatches !== right.metadataMatches) return right.metadataMatches - left.metadataMatches;
-    return compareRecencyDesc(left.record, right.record);
-  });
-  return scored.map((item) => item.record);
-}
-
-function selectRecords(methodId, records, query, topK) {
-  if (methodId === 'full_context') return records;
-  if (methodId === 'recency_last_n') return [...records].sort(compareRecencyDesc).slice(0, topK);
-  if (methodId === 'keyword_filter') return rankedByOverlap(records, query.question).slice(0, topK);
-  if (methodId === 'enigma_relevance') return rankedByEnigmaRelevance(records, query).slice(0, topK);
-  throw new Error(`Unknown method ${methodId}`);
-}
-
-function estimatePromptTokens(question, selectedRecords) {
-  let tokens = estimateTextTokens(question);
-  for (const record of selectedRecords) tokens += record.estimated_tokens;
-  return tokens;
-}
-
-function percentile(values, ratio) {
-  if (values.length === 0) return 0;
-  const sorted = [...values].sort((left, right) => left - right);
-  const index = Math.min(sorted.length - 1, Math.max(0, Math.ceil(sorted.length * ratio) - 1));
-  return sorted[index];
-}
-
-function latencySummary(values) {
-  return {
-    samples: values.length,
-    p50_ms: round6(percentile(values, 0.5)),
-    p95_ms: round6(percentile(values, 0.95)),
-    min_ms: round6(values.length === 0 ? 0 : Math.min(...values)),
-    max_ms: round6(values.length === 0 ? 0 : Math.max(...values)),
-  };
-}
-
-function rate(numerator, denominator) {
-  if (denominator === 0) return null;
-  return round6(numerator / denominator);
-}
-
-function round6(value) {
-  return Number(value.toFixed(6));
-}
-
-function mean(total, count) {
-  return count === 0 ? 0 : round6(total / count);
-}
-
-function recordFromContent(args) {
-  const content = String(args.content ?? '');
-  return {
-    id: args.id,
-    dataset_item_id: args.dataset_item_id,
-    session_id: args.session_id,
-    turn_id: args.turn_id,
-    dialog_id: args.dialog_id,
-    role: args.role,
-    kind: args.kind,
-    tags: args.tags ?? [],
-    has_answer: args.has_answer === true,
-    ordinal: args.ordinal,
-    content,
-    estimated_tokens: estimateTextTokens(content),
-  };
-}
-
-function parseLocomoEvidenceLabels(evidence) {
-  const labels = new Set();
-  const stack = Array.isArray(evidence) ? [...evidence] : [evidence];
-  while (stack.length > 0) {
-    const value = stack.shift();
-    if (Array.isArray(value)) {
-      stack.push(...value);
-      continue;
-    }
-    if (value === undefined || value === null) continue;
-    for (const match of String(value).matchAll(/D\s*(\d+)\s*:\s*(\d+)/giu)) {
-      labels.add(`D${Number(match[1])}:${Number(match[2])}`);
-    }
-  }
-  return labels;
-}
-
-function dialogIdForTurn(sessionNumber, turn, turnIndex) {
-  const raw = turn?.dia_id ?? turn?.dialog_id ?? turn?.turn_id ?? turn?.id ?? turnIndex + 1;
-  const text = String(raw);
-  const match = text.match(/^D\s*(\d+)\s*:\s*(\d+)$/iu);
-  if (match) return `D${Number(match[1])}:${Number(match[2])}`;
-  const numeric = text.match(/\d+/u)?.[0] ?? String(turnIndex + 1);
-  return `D${sessionNumber}:${Number(numeric)}`;
-}
-
-export function parseLocomoDataset(data, options = {}) {
-  const maxQa = optionalPositiveInteger(options.max_qa ?? options.maxQa, 'max_locomo_qa');
-  const rows = normalizeDatasetArray(data, 'LoCoMo');
-  const records = [];
-  const queries = [];
-  let ordinal = 0;
-  for (let sampleIndex = 0; sampleIndex < rows.length; sampleIndex += 1) {
-    const sample = rows[sampleIndex] ?? {};
-    const itemId = String(sample.sample_id ?? sample.id ?? `sample_${sampleIndex + 1}`);
-    const conversation = sample.conversation ?? {};
-    const sessionNames = Object.keys(conversation)
-      .map((key) => {
-        const match = key.match(/^session_(\d+)$/u);
-        return match ? { key, sessionNumber: Number(match[1]) } : null;
-      })
-      .filter(Boolean)
-      .sort((left, right) => left.sessionNumber - right.sessionNumber);
-    for (const { key, sessionNumber } of sessionNames) {
-      const session = conversation[key];
-      if (!Array.isArray(session)) continue;
-      for (let turnIndex = 0; turnIndex < session.length; turnIndex += 1) {
-        const turn = session[turnIndex] ?? {};
-        const dialogId = dialogIdForTurn(sessionNumber, turn, turnIndex);
-        records.push(recordFromContent({
-          id: `locomo:${itemId}:${dialogId}`,
-          dataset_item_id: itemId,
-          session_id: `D${sessionNumber}`,
-          turn_id: dialogId,
-          dialog_id: dialogId,
-          role: turn.speaker ?? turn.role ?? undefined,
-          kind: 'locomo_conversation_turn',
-          tags: ['locomo', `session_${sessionNumber}`],
-          ordinal,
-          content: turn.text ?? turn.content ?? turn.message ?? '',
-        }));
-        ordinal += 1;
-      }
-    }
-    const qaRows = Array.isArray(sample.qa) ? sample.qa : [];
-    for (let qaIndex = 0; qaIndex < qaRows.length; qaIndex += 1) {
-      if (maxQa !== undefined && queries.length >= maxQa) break;
-      const qa = qaRows[qaIndex] ?? {};
-      const evidenceDialogIds = parseLocomoEvidenceLabels(qa.evidence);
-      queries.push({
-        id: `locomo:${itemId}:qa_${qaIndex + 1}`,
-        dataset_item_id: itemId,
-        question: String(qa.question ?? ''),
-        category: qa.category === undefined ? undefined : String(qa.category),
-        evidence_dialog_ids: evidenceDialogIds,
-        abstention: evidenceDialogIds.size === 0,
-      });
-    }
-    if (maxQa !== undefined && queries.length >= maxQa) break;
-  }
-  return {
-    id: 'locomo',
-    label: 'LoCoMo',
-    source_url: LOCOMO_SOURCE_URL,
-    license: 'CC BY-NC 4.0',
-    parser: 'conversation session turns as memory records; qa evidence labels mapped to dialog ids such as D1:3 and semicolon-separated labels',
-    task_categories: ['multi-session QA', 'event summarization', 'multimodal generation over long conversations'],
-    records,
-    queries,
-  };
-}
-
-function sessionIdAt(ids, index) {
-  if (Array.isArray(ids) && ids[index] !== undefined && ids[index] !== null) return String(ids[index]);
-  return String(index);
-}
-
-function answerSessionIds(item) {
-  if (!Array.isArray(item.answer_session_ids)) return new Set();
-  return new Set(item.answer_session_ids.map((id) => String(id)));
-}
-
-function isLongMemEvalAbstention(item, answerSessions) {
-  const id = String(item.question_id ?? item.id ?? '');
-  if (id.endsWith('_abs') || id.includes('_abs_')) return true;
-  if (String(item.question_type ?? '').toLowerCase().includes('abst')) return true;
-  return answerSessions.size === 0;
-}
-
-export function parseLongMemEvalDataset(data, options = {}) {
-  const maxItems = optionalPositiveInteger(options.max_items ?? options.maxItems, 'max_longmemeval_items');
-  const rows = normalizeDatasetArray(data, 'LongMemEval');
-  const records = [];
-  const queries = [];
-  let ordinal = 0;
-  const limit = maxItems === undefined ? rows.length : Math.min(rows.length, maxItems);
-  for (let itemIndex = 0; itemIndex < limit; itemIndex += 1) {
-    const item = rows[itemIndex] ?? {};
-    const itemId = String(item.question_id ?? item.id ?? `item_${itemIndex + 1}`);
-    const sessions = Array.isArray(item.haystack_sessions) ? item.haystack_sessions : [];
-    const sessionIds = item.haystack_session_ids;
-    const evidenceTurnIds = new Set();
-    for (let sessionIndex = 0; sessionIndex < sessions.length; sessionIndex += 1) {
-      const session = sessions[sessionIndex];
-      if (!Array.isArray(session)) continue;
-      const sessionId = sessionIdAt(sessionIds, sessionIndex);
-      for (let turnIndex = 0; turnIndex < session.length; turnIndex += 1) {
-        const turn = session[turnIndex] ?? {};
-        const turnId = `${sessionId}:${turnIndex}`;
-        if (turn.has_answer === true) evidenceTurnIds.add(turnId);
-        records.push(recordFromContent({
-          id: `longmemeval:${itemId}:${turnId}`,
-          dataset_item_id: itemId,
-          session_id: sessionId,
-          turn_id: turnId,
-          role: turn.role ?? undefined,
-          kind: 'longmemeval_haystack_turn',
-          tags: ['longmemeval', String(item.question_type ?? ''), `session_${sessionId}`],
-          has_answer: turn.has_answer === true,
-          ordinal,
-          content: turn.content ?? turn.text ?? turn.message ?? '',
-        }));
-        ordinal += 1;
-      }
-    }
-    const answerSessions = answerSessionIds(item);
-    queries.push({
-      id: `longmemeval:${itemId}`,
-      dataset_item_id: itemId,
-      question: String(item.question ?? ''),
-      question_type: item.question_type === undefined ? undefined : String(item.question_type),
-      evidence_turn_ids: evidenceTurnIds,
-      evidence_session_ids: answerSessions,
-      abstention: isLongMemEvalAbstention(item, answerSessions),
-    });
-  }
-  return {
-    id: 'longmemeval',
-    label: 'LongMemEval',
-    source_url: LONGMEMEVAL_SOURCE_URLS,
-    license: 'See Hugging Face dataset card and upstream LongMemEval repository for the selected cleaned file.',
-    parser: 'haystack_sessions turns as memory records; has_answer:true turns and answer_session_ids are used as evidence labels; _abs ids are evaluated as abstention cases',
-    task_categories: ['information extraction', 'multi-session reasoning', 'temporal reasoning', 'knowledge updates', 'abstention'],
-    records,
-    queries,
-  };
-}
-
-function scoreLocomoMethod(method, dataset, topK) {
-  const latencies = [];
-  let evidenceQuestions = 0;
-  let hits = 0;
-  let exactCoverage = 0;
-  let totalTokens = 0;
-  let selectedTotal = 0;
-  for (const query of dataset.queries) {
-    const start = performance.now();
-    const records = dataset.records.filter((record) => record.dataset_item_id === query.dataset_item_id);
-    const selected = selectRecords(method.id, records, query, topK);
-    latencies.push(performance.now() - start);
-    const selectedDialogs = new Set(selected.map((record) => record.dialog_id));
-    selectedTotal += selected.length;
-    totalTokens += estimatePromptTokens(query.question, selected);
-    if (query.evidence_dialog_ids.size > 0) {
-      evidenceQuestions += 1;
-      let covered = 0;
-      for (const evidenceId of query.evidence_dialog_ids) if (selectedDialogs.has(evidenceId)) covered += 1;
-      if (covered > 0) hits += 1;
-      if (covered === query.evidence_dialog_ids.size) exactCoverage += 1;
-    }
-  }
-  return {
-    id: method.id,
-    method: method.id,
-    local_method_only: true,
-    external_provider_called: false,
-    retrieval_proxy_only: true,
-    uses_top_k: method.uses_top_k,
-    top_k: method.uses_top_k ? topK : null,
-    question_count: dataset.queries.length,
-    evidence_question_count: evidenceQuestions,
-    evidence_hit_at_k: rate(hits, evidenceQuestions),
-    exact_evidence_coverage: rate(exactCoverage, evidenceQuestions),
-    estimated_prompt_tokens: {
-      total: totalTokens,
-      mean_per_question: mean(totalTokens, dataset.queries.length),
-      estimator: 'estimateTextTokens deterministic local estimator',
-    },
-    selected_memory_count: {
-      total: selectedTotal,
-      mean_per_question: mean(selectedTotal, dataset.queries.length),
-    },
-    latency: latencySummary(latencies),
-    public_question_text_included: false,
-    public_answer_text_included: false,
-    raw_conversation_text_included: false,
-  };
-}
-
-function scoreLongMemEvalMethod(method, dataset, topK) {
-  const latencies = [];
-  let turnEvidenceQuestions = 0;
-  let turnHits = 0;
-  let exactTurnCoverage = 0;
-  let sessionEvidenceQuestions = 0;
-  let sessionHits = 0;
-  let exactSessionCoverage = 0;
-  let abstentionQuestions = 0;
-  let abstentionCorrect = 0;
-  let totalTokens = 0;
-  let selectedTotal = 0;
-  for (const query of dataset.queries) {
-    const start = performance.now();
-    const records = dataset.records.filter((record) => record.dataset_item_id === query.dataset_item_id);
-    const selected = selectRecords(method.id, records, query, topK);
-    latencies.push(performance.now() - start);
-    const selectedTurns = new Set(selected.map((record) => record.turn_id));
-    const selectedSessions = new Set(selected.map((record) => record.session_id));
-    selectedTotal += selected.length;
-    totalTokens += estimatePromptTokens(query.question, selected);
-
-    if (query.abstention) {
-      abstentionQuestions += 1;
-      let selectedGold = false;
-      for (const turnId of query.evidence_turn_ids) if (selectedTurns.has(turnId)) selectedGold = true;
-      for (const sessionId of query.evidence_session_ids) if (selectedSessions.has(sessionId)) selectedGold = true;
-      if (!selectedGold) abstentionCorrect += 1;
-      continue;
-    }
-
-    if (query.evidence_turn_ids.size > 0) {
-      turnEvidenceQuestions += 1;
-      let covered = 0;
-      for (const turnId of query.evidence_turn_ids) if (selectedTurns.has(turnId)) covered += 1;
-      if (covered > 0) turnHits += 1;
-      if (covered === query.evidence_turn_ids.size) exactTurnCoverage += 1;
-    }
-    if (query.evidence_session_ids.size > 0) {
-      sessionEvidenceQuestions += 1;
-      let covered = 0;
-      for (const sessionId of query.evidence_session_ids) if (selectedSessions.has(sessionId)) covered += 1;
-      if (covered > 0) sessionHits += 1;
-      if (covered === query.evidence_session_ids.size) exactSessionCoverage += 1;
-    }
-  }
-  return {
-    id: method.id,
-    method: method.id,
-    local_method_only: true,
-    external_provider_called: false,
-    retrieval_proxy_only: true,
-    uses_top_k: method.uses_top_k,
-    top_k: method.uses_top_k ? topK : null,
-    item_count: dataset.queries.length,
-    turn_evidence_question_count: turnEvidenceQuestions,
-    turn_evidence_hit_at_k: rate(turnHits, turnEvidenceQuestions),
-    exact_turn_evidence_coverage: rate(exactTurnCoverage, turnEvidenceQuestions),
-    session_evidence_question_count: sessionEvidenceQuestions,
-    session_evidence_hit_at_k: rate(sessionHits, sessionEvidenceQuestions),
-    exact_session_evidence_coverage: rate(exactSessionCoverage, sessionEvidenceQuestions),
-    abstention_questions: abstentionQuestions,
-    abstention_correct: abstentionCorrect,
-    abstention_correctness: rate(abstentionCorrect, abstentionQuestions),
-    estimated_prompt_tokens: {
-      total: totalTokens,
-      mean_per_item: mean(totalTokens, dataset.queries.length),
-      estimator: 'estimateTextTokens deterministic local estimator',
-    },
-    selected_memory_count: {
-      total: selectedTotal,
-      mean_per_item: mean(selectedTotal, dataset.queries.length),
-    },
-    latency: latencySummary(latencies),
-    public_question_text_included: false,
-    public_answer_text_included: false,
-    raw_conversation_text_included: false,
-  };
-}
-
-function scoreDataset(dataset, topK) {
-  const methodRows = STANDARD_MEMORY_BENCHMARK_METHODS.map((method) => (
-    dataset.id === 'locomo' ? scoreLocomoMethod(method, dataset, topK) : scoreLongMemEvalMethod(method, dataset, topK)
-  ));
-  return {
-    id: dataset.id,
-    dataset: dataset.id,
-    label: dataset.label,
-    source_url: dataset.source_url,
-    license: dataset.license,
-    parser: dataset.parser,
-    task_categories: dataset.task_categories,
-    record_count: dataset.records.length,
-    question_count: dataset.queries.length,
-    item_count: dataset.queries.length,
-    raw_question_text_included: false,
-    raw_answer_text_included: false,
-    raw_conversation_text_included: false,
-    methods: methodRows,
-  };
-}
-
-export function runStandardMemoryBenchmarkSuite(options = {}) {
-  const topK = optionalPositiveInteger(options.top_k ?? options.topK, 'top_k') ?? 5;
-  const datasetRows = [];
-  if (options.locomoData !== undefined) {
-    datasetRows.push(scoreDataset(parseLocomoDataset(options.locomoData, { max_qa: options.max_locomo_qa ?? options.maxLocomoQa }), topK));
-  }
-  if (options.longMemEvalData !== undefined || options.longmemevalData !== undefined) {
-    datasetRows.push(scoreDataset(parseLongMemEvalDataset(options.longMemEvalData ?? options.longmemevalData, { max_items: options.max_longmemeval_items ?? options.maxLongMemEvalItems }), topK));
-  }
-  if (datasetRows.length === 0) throw new Error('At least one standard dataset is required: provide locomoData and/or longMemEvalData');
-  return buildSuiteReport(datasetRows, topK, options);
-}
-
-export async function runStandardMemoryBenchmarkSuiteFromFiles(options = {}) {
-  const datasetRows = [];
-  const topK = optionalPositiveInteger(options.top_k ?? options.topK, 'top_k') ?? 5;
-  if (options.locomo !== undefined || options.locomoPath !== undefined) {
-    const path = options.locomo ?? options.locomoPath;
-    const loaded = await readJsonWithSha256(path);
-    datasetRows.push({
-      ...scoreDataset(parseLocomoDataset(loaded.data, { max_qa: options.max_locomo_qa ?? options.maxLocomoQa }), topK),
-      local_file_name: publicFileName(path),
-      input_sha256: loaded.sha256,
-    });
-  }
-  if (options.longmemeval !== undefined || options.longmemevalPath !== undefined || options.longMemEvalPath !== undefined) {
-    const path = options.longmemeval ?? options.longmemevalPath ?? options.longMemEvalPath;
-    const maxItems = optionalPositiveInteger(options.max_longmemeval_items ?? options.maxLongMemEvalItems, 'max_longmemeval_items');
-    const loaded = await readLongMemEvalJsonWithSha256(path, maxItems);
-    datasetRows.push({
-      ...scoreDataset(parseLongMemEvalDataset(loaded.data, { max_items: maxItems }), topK),
-      local_file_name: publicFileName(path),
-      input_sha256: loaded.sha256,
-    });
-  }
-  if (datasetRows.length === 0) throw new Error('Provide --locomo <path> and/or --longmemeval <path>');
-  return buildSuiteReport(datasetRows, topK, options);
-}
-
-function buildSuiteReport(datasetRows, topK, options) {
-  return {
-    schema: STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA,
-    generated_at: options.generated_at ?? new Date().toISOString(),
-    package: {
-      name: 'enigma-memory',
-      version: '0.1.13',
-    },
-    public_safe: true,
-    top_k: topK,
-    source_urls: {
-      locomo: LOCOMO_SOURCE_URL,
-      longmemeval: LONGMEMEVAL_SOURCE_URLS,
-    },
-    license_and_boundary_notes: [
-      'LoCoMo source data is CC BY-NC 4.0; keep local dataset files and raw conversations out of public reports unless separately reviewed.',
-      'LongMemEval cleaned files are operator-supplied local JSON files from the upstream Hugging Face dataset repository.',
-      'Scores are retrieval/evidence proxy metrics over official dataset labels, not LLM-generated answer accuracy.',
-      'No provider APIs, hosted runtimes, competitor SDKs, or external accounts are called by this runner.',
-      'Rows are local deterministic methods only; no third-party competitor scores or benchmark-leadership claims are emitted.',
-    ],
-    benchmark_boundaries: {
-      official_dataset_files_required: true,
-      credentials_required: false,
-      external_provider_calls: false,
-      llm_answer_accuracy_scored: false,
-      retrieval_evidence_proxy_scored: true,
-      raw_question_text_included: false,
-      raw_answer_text_included: false,
-      raw_conversation_text_included: false,
-      provider_deletion_claim: false,
-      model_forgetting_claim: false,
-      roi_or_provider_invoice_savings_claim: false,
-      compliance_certification_claim: false,
-      benchmark_leadership_claim: false,
-    },
-    relevance_logic: {
-      token_extraction: 'keyword_filter uses basic lowercase /[a-z0-9]+(?:[-_][a-z0-9]+)*/ overlap after stopword removal; enigma_relevance additionally applies deterministic suffix stemming for ing, ed, and plural s forms',
-      production_alignment: 'public-safe local approximation of production query-aware retrieval using query/content stems, role/session/kind/tag hints, category/task hints, temporal/date hints, and phrase/proximity boosts; no private memory is emitted',
-      keyword_filter_fallback: 'empty result when no query/content token overlap exists',
-      enigma_relevance_fallback: 'falls back to all local candidates only when no enhanced relevance signal exists, then applies deterministic local ranking and --top-k',
-      provider_api_used: false,
-      llm_used: false,
-    },
-    local_methods: STANDARD_MEMORY_BENCHMARK_METHODS.map((method) => ({ ...method })),
-    datasets: datasetRows,
-    dataset_rows: datasetRows,
-  };
-}
-
-function usage() {
-  return `Usage: node scripts/run-standard-memory-benchmarks.mjs [--locomo <path>] [--longmemeval <path>] [--max-locomo-qa <n>] [--max-longmemeval-items <n>] [--top-k <n>] [--out <path>]\n\nProduces schema ${STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA}. Raw question, answer, and conversation text are never written to the report. With --longmemeval and --max-longmemeval-items, the local top-level JSON array is streamed for hashing and only the requested sample items are parsed.`;
-}
-
-async function main() {
-  const options = parseArgs();
-  if (options.help) {
-    console.log(usage());
-    return;
-  }
-  const report = await runStandardMemoryBenchmarkSuiteFromFiles(options);
-  const serialized = `${JSON.stringify(report, null, 2)}\n`;
-  if (options.out) {
-    const outPath = resolve(options.out);
-    await mkdir(dirname(outPath), { recursive: true });
-    await writeFile(outPath, serialized);
-  } else {
-    process.stdout.write(serialized);
-  }
-}
-
-const invokedPath = process.argv[1] ? resolve(process.argv[1]) : '';
-const modulePath = fileURLToPath(import.meta.url);
-if (invokedPath === modulePath) {
-  main().catch((error) => {
-    console.error(error instanceof Error ? error.message : String(error));
-    process.exitCode = 1;
-  });
-}
+#!/usr/bin/env node
+import { createReadStream } from 'node:fs';
+import { mkdir, readFile, writeFile } from 'node:fs/promises';
+import { basename, dirname, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { createHash } from 'node:crypto';
+import { performance } from 'node:perf_hooks';
+import { StringDecoder } from 'node:string_decoder';
+import { estimateTextTokens } from '../packages/optimizer/src/index.js';
+
+export const STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA = 'enigma.standard_memory_benchmark_suite.v1';
+
+export const STANDARD_MEMORY_BENCHMARK_METHODS = Object.freeze([
+  Object.freeze({
+    id: 'full_context',
+    label: 'Full context',
+    boundary: 'Supplies every parsed memory record for each query; no provider API or model answer generation.',
+    uses_top_k: false,
+  }),
+  Object.freeze({
+    id: 'recency_last_n',
+    label: 'Recency last N',
+    boundary: 'Supplies the most recent local memory records up to --top-k.',
+    uses_top_k: true,
+  }),
+  Object.freeze({
+    id: 'keyword_filter',
+    label: 'Keyword filter',
+    boundary: 'Supplies local memory records whose public-safe deterministic tokens overlap the query, capped by --top-k.',
+    uses_top_k: true,
+  }),
+  Object.freeze({
+    id: 'enigma_relevance',
+    label: 'Enigma relevance',
+    boundary: 'Uses deterministic query-aware relevance features over local public-safe memory metadata and content tokens, then ranks locally without provider APIs.',
+    uses_top_k: true,
+  }),
+]);
+
+const LOCOMO_SOURCE_URL = 'https://raw.githubusercontent.com/snap-research/locomo/main/data/locomo10.json';
+const LONGMEMEVAL_SOURCE_URLS = Object.freeze([
+  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_oracle.json',
+  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json',
+  'https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_m_cleaned.json',
+]);
+
+const QUERY_RELEVANCE_STOPWORDS = new Set([
+  'about',
+  'after',
+  'again',
+  'against',
+  'also',
+  'and',
+  'any',
+  'are',
+  'assistant',
+  'because',
+  'been',
+  'before',
+  'being',
+  'between',
+  'can',
+  'could',
+  'current',
+  'does',
+  'from',
+  'has',
+  'have',
+  'how',
+  'into',
+  'its',
+  'latest',
+  'more',
+  'most',
+  'number',
+  'own',
+  'owns',
+  'please',
+  'should',
+  'that',
+  'the',
+  'their',
+  'then',
+  'there',
+  'these',
+  'they',
+  'this',
+  'use',
+  'using',
+  'was',
+  'what',
+  'when',
+  'where',
+  'which',
+  'who',
+  'whose',
+  'why',
+  'with',
+  'would',
+]);
+
+function parseArgs(argv = process.argv.slice(2)) {
+  const options = { top_k: 5 };
+  for (let index = 0; index < argv.length; index += 1) {
+    const arg = argv[index];
+    if (arg === '--locomo') {
+      options.locomo = requiredFlagValue(argv, index, arg);
+      index += 1;
+    } else if (arg === '--longmemeval') {
+      options.longmemeval = requiredFlagValue(argv, index, arg);
+      index += 1;
+    } else if (arg === '--max-locomo-qa') {
+      options.max_locomo_qa = positiveInteger(requiredFlagValue(argv, index, arg), arg);
+      index += 1;
+    } else if (arg === '--max-longmemeval-items') {
+      options.max_longmemeval_items = positiveInteger(requiredFlagValue(argv, index, arg), arg);
+      index += 1;
+    } else if (arg === '--top-k') {
+      options.top_k = positiveInteger(requiredFlagValue(argv, index, arg), arg);
+      index += 1;
+    } else if (arg === '--out') {
+      options.out = requiredFlagValue(argv, index, arg);
+      index += 1;
+    } else if (arg === '--help' || arg === '-h') {
+      options.help = true;
+    } else {
+      throw new Error(`Unknown option ${arg}`);
+    }
+  }
+  return options;
+}
+
+function requiredFlagValue(argv, index, flag) {
+  const value = argv[index + 1];
+  if (value === undefined || value.startsWith('--')) throw new Error(`${flag} requires a value`);
+  return value;
+}
+
+function positiveInteger(value, name) {
+  const number = Number(value);
+  if (!Number.isInteger(number) || number <= 0) throw new Error(`${name} must be a positive integer`);
+  return number;
+}
+
+function optionalPositiveInteger(value, name) {
+  if (value === undefined || value === null) return undefined;
+  return positiveInteger(value, name);
+}
+
+function publicFileName(path) {
+  return path === undefined || path === null ? undefined : basename(String(path));
+}
+
+async function readJsonWithSha256(path) {
+  const raw = await readFile(path, 'utf8');
+  return {
+    data: JSON.parse(raw),
+    sha256: createHash('sha256').update(raw).digest('hex'),
+  };
+}
+
+function isJsonWhitespace(char) {
+  const code = char.charCodeAt(0);
+  return code === 0x20 || code === 0x0a || code === 0x0d || code === 0x09 || code === 0xfeff;
+}
+
+function createTopLevelArraySampler(maxItems, name) {
+  const items = [];
+  let started = false;
+  let closed = false;
+  let collecting = false;
+  let doneCollecting = false;
+  let expectSeparator = false;
+  let depth = 0;
+  let inString = false;
+  let escaped = false;
+  let current = '';
+
+  function fail(message) {
+    throw new Error(`${name} sample-mode JSON parse failed: ${message}`);
+  }
+
+  function finishItem() {
+    items.push(current);
+    current = '';
+    collecting = false;
+    expectSeparator = true;
+    if (items.length >= maxItems) doneCollecting = true;
+  }
+
+  return {
+    write(text) {
+      for (const char of text) {
+        if (!started) {
+          if (isJsonWhitespace(char)) continue;
+          if (char !== '[') fail('expected a top-level array');
+          started = true;
+          continue;
+        }
+
+        if (doneCollecting) continue;
+
+        if (closed) {
+          if (!isJsonWhitespace(char)) fail('found trailing data after the top-level array');
+          continue;
+        }
+
+        if (!collecting) {
+          if (isJsonWhitespace(char)) continue;
+          if (expectSeparator) {
+            if (char === ',') {
+              expectSeparator = false;
+              continue;
+            }
+            if (char === ']') {
+              closed = true;
+              continue;
+            }
+            fail('expected a comma or closing bracket between items');
+          }
+          if (char === ']') {
+            closed = true;
+            continue;
+          }
+          if (char !== '{' && char !== '[') fail('expected each sampled item to be an object or array');
+          collecting = true;
+          current = char;
+          depth = 1;
+          continue;
+        }
+
+        current += char;
+        if (inString) {
+          if (escaped) {
+            escaped = false;
+          } else if (char === '\\') {
+            escaped = true;
+          } else if (char === '"') {
+            inString = false;
+          }
+          continue;
+        }
+        if (char === '"') {
+          inString = true;
+        } else if (char === '{' || char === '[') {
+          depth += 1;
+        } else if (char === '}' || char === ']') {
+          depth -= 1;
+          if (depth < 0) fail('encountered an unmatched closing bracket');
+          if (depth === 0) finishItem();
+        }
+      }
+    },
+    get done() {
+      return doneCollecting;
+    },
+    finish() {
+      if (!started) fail('empty input');
+      if (!doneCollecting && collecting) fail('ended inside a sampled item');
+      if (!doneCollecting && inString) fail('ended inside a string');
+      if (!doneCollecting && !closed) fail('ended before the top-level array closed');
+      return JSON.parse(`[${items.join(',')}]`);
+    },
+  };
+}
+
+async function readJsonArraySampleWithSha256(path, maxItems, name) {
+  const hash = createHash('sha256');
+  const decoder = new StringDecoder('utf8');
+  const sampler = createTopLevelArraySampler(maxItems, name);
+  for await (const chunk of createReadStream(path)) {
+    hash.update(chunk);
+    if (!sampler.done) sampler.write(decoder.write(chunk));
+  }
+  if (!sampler.done) {
+    const tail = decoder.end();
+    if (tail.length > 0) sampler.write(tail);
+  }
+  return {
+    data: sampler.finish(),
+    sha256: hash.digest('hex'),
+  };
+}
+
+async function readLongMemEvalJsonWithSha256(path, maxItems) {
+  if (maxItems === undefined) return readJsonWithSha256(path);
+  try {
+    return await readJsonArraySampleWithSha256(path, maxItems, 'LongMemEval');
+  } catch (error) {
+    if (error instanceof Error && error.message === 'LongMemEval sample-mode JSON parse failed: expected a top-level array') {
+      return readJsonWithSha256(path);
+    }
+    throw error;
+  }
+}
+
+function normalizeDatasetArray(data, name) {
+  if (Array.isArray(data)) return data;
+  if (data && typeof data === 'object') {
+    for (const key of ['data', 'items', 'examples', 'samples']) {
+      if (Array.isArray(data[key])) return data[key];
+    }
+  }
+  throw new TypeError(`${name} dataset must be a JSON array or object containing an array`);
+}
+
+function addMeaningfulToken(tokens, token) {
+  if (token.length < 3) return;
+  if (!/[a-z]/u.test(token)) return;
+  if (QUERY_RELEVANCE_STOPWORDS.has(token)) return;
+  tokens.add(token);
+}
+
+function stemToken(token) {
+  if (token.length > 5 && token.endsWith('ing')) {
+    let stem = token.slice(0, -3);
+    if (stem.length > 3 && stem.at(-1) === stem.at(-2)) stem = stem.slice(0, -1);
+    return stem;
+  }
+  if (token.length > 4 && token.endsWith('ed')) {
+    let stem = token.slice(0, -2);
+    if (stem.length > 3 && stem.at(-1) === stem.at(-2)) stem = stem.slice(0, -1);
+    return stem;
+  }
+  if (token.length > 4 && token.endsWith('ies')) return `${token.slice(0, -3)}y`;
+  if (token.length > 4 && token.endsWith('es')) return token.slice(0, -2);
+  if (token.length > 3 && token.endsWith('s') && !token.endsWith('ss')) return token.slice(0, -1);
+  return token;
+}
+
+function addStemmedMeaningfulToken(tokens, token) {
+  addMeaningfulToken(tokens, token);
+  const stem = stemToken(token);
+  addMeaningfulToken(tokens, stem);
+  if (token.endsWith('ed') || token.endsWith('ing')) addMeaningfulToken(tokens, `${stem}e`);
+}
+
+function meaningfulTokensFrom(value) {
+  const tokens = new Set();
+  if (value === undefined || value === null) return tokens;
+  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
+    const token = match[0];
+    addMeaningfulToken(tokens, token);
+    if (token.includes('-') || token.includes('_')) {
+      for (const part of token.split(/[-_]+/u)) addMeaningfulToken(tokens, part);
+    }
+  }
+  return tokens;
+}
+
+function stemmedMeaningfulTokensFrom(value) {
+  const tokens = new Set();
+  if (value === undefined || value === null) return tokens;
+  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
+    const token = match[0];
+    addStemmedMeaningfulToken(tokens, token);
+    if (token.includes('-') || token.includes('_')) {
+      for (const part of token.split(/[-_]+/u)) addStemmedMeaningfulToken(tokens, part);
+    }
+  }
+  return tokens;
+}
+
+function stemmedTokenSequenceFrom(value) {
+  const sequence = [];
+  if (value === undefined || value === null) return sequence;
+  for (const match of String(value).toLowerCase().matchAll(/[a-z0-9]+(?:[-_][a-z0-9]+)*/gu)) {
+    const token = match[0];
+    const parts = token.includes('-') || token.includes('_') ? token.split(/[-_]+/u) : [token];
+    for (const part of parts) {
+      const stem = stemToken(part);
+      if (stem.length >= 3 && /[a-z]/u.test(stem) && !QUERY_RELEVANCE_STOPWORDS.has(stem)) sequence.push(stem);
+    }
+  }
+  return sequence;
+}
+
+function tokenOverlapScore(queryTokens, record) {
+  if (queryTokens.size === 0) return 0;
+  const recordTokens = meaningfulTokensFrom(record.content);
+  let score = 0;
+  for (const token of queryTokens) if (recordTokens.has(token)) score += 1;
+  return score;
+}
+
+function normalizedTagSessionToken(value) {
+  return String(value ?? '').toLowerCase().replace(/^session[-_:]?/u, '').replace(/[^a-z0-9]+/gu, '');
+}
+
+function roleHintsFrom(value) {
+  const hints = new Set();
+  const text = String(value ?? '').toLowerCase();
+  for (const match of text.matchAll(/\b(?:assistant|user|system|human|agent|speaker[-_\s]?[a-z0-9]+)\b/gu)) {
+    const compact = match[0].replace(/\s+/gu, '_');
+    hints.add(compact);
+    for (const token of stemmedMeaningfulTokensFrom(compact)) hints.add(token);
+  }
+  return hints;
+}
+
+function sessionHintsFrom(value) {
+  const hints = new Set();
+  const text = String(value ?? '').toLowerCase();
+  for (const match of text.matchAll(/\b(?:session|sess)\s*[-_:]?\s*([a-z0-9]+)\b/gu)) hints.add(match[1]);
+  for (const match of text.matchAll(/\bd\s*[-_:]?\s*(\d+)\b/gu)) hints.add(`d${Number(match[1])}`);
+  for (const match of text.matchAll(/\bsession[-_]([a-z0-9]+)\b/gu)) hints.add(match[1]);
+  return hints;
+}
+
+function dateHintsFrom(value) {
+  const hints = new Set();
+  const text = String(value ?? '').toLowerCase();
+  for (const match of text.matchAll(/\b(?:19|20)\d{2}\b/gu)) hints.add(match[0]);
+  for (const match of text.matchAll(/\b\d{4}[-/]\d{1,2}(?:[-/]\d{1,2})?\b/gu)) hints.add(match[0].replace(/\D+/gu, '-'));
+  for (const match of text.matchAll(/\b(?:jan(?:uary)?|feb(?:ruary)?|mar(?:ch)?|apr(?:il)?|may|jun(?:e)?|jul(?:y)?|aug(?:ust)?|sep(?:t(?:ember)?)?|oct(?:ober)?|nov(?:ember)?|dec(?:ember)?)\b/gu)) hints.add(match[0].slice(0, 3));
+  for (const match of text.matchAll(/\b(?:today|yesterday|tomorrow|recent|recently|latest|newest|current|previous|last|earliest|oldest)\b/gu)) hints.add(match[0]);
+  return hints;
+}
+
+function recordSessionTokens(record) {
+  const tokens = new Set();
+  const values = [record.session_id, record.dialog_id, record.turn_id, ...(record.tags ?? [])];
+  for (const value of values) {
+    if (value === undefined || value === null) continue;
+    const normalized = normalizedTagSessionToken(value);
+    if (normalized) tokens.add(normalized);
+    for (const hint of sessionHintsFrom(value)) tokens.add(normalizedTagSessionToken(hint));
+  }
+  return tokens;
+}
+
+function countSetIntersection(left, right) {
+  let count = 0;
+  for (const value of left) if (right.has(value)) count += 1;
+  return count;
+}
+
+function phraseAndProximityScore(querySequence, recordSequence) {
+  if (querySequence.length < 2 || recordSequence.length < 2) return 0;
+  const recordBigrams = new Set();
+  const positions = new Map();
+  for (let index = 0; index < recordSequence.length; index += 1) {
+    const token = recordSequence[index];
+    if (!positions.has(token)) positions.set(token, []);
+    positions.get(token).push(index);
+    if (index > 0) recordBigrams.add(`${recordSequence[index - 1]}\u0000${token}`);
+  }
+  let score = 0;
+  for (let index = 1; index < querySequence.length; index += 1) {
+    const previous = querySequence[index - 1];
+    const current = querySequence[index];
+    if (previous === current) continue;
+    if (recordBigrams.has(`${previous}\u0000${current}`)) {
+      score += 10;
+      continue;
+    }
+    const leftPositions = positions.get(previous);
+    const rightPositions = positions.get(current);
+    if (!leftPositions || !rightPositions) continue;
+    let near = false;
+    for (const left of leftPositions) {
+      for (const right of rightPositions) {
+        if (Math.abs(left - right) <= 6) {
+          near = true;
+          break;
+        }
+      }
+      if (near) break;
+    }
+    if (near) score += 4;
+  }
+  return score;
+}
+
+function hasRecencyIntent(hints) {
+  for (const hint of hints) {
+    if (hint === 'recent' || hint === 'recently' || hint === 'latest' || hint === 'newest' || hint === 'current' || hint === 'previous' || hint === 'last') return true;
+  }
+  return false;
+}
+
+function enigmaRelevanceScore(query, record) {
+  const question = String(query.question ?? '');
+  const queryTokens = stemmedMeaningfulTokensFrom(question);
+  const categoryTokens = stemmedMeaningfulTokensFrom(`${query.category ?? ''} ${query.question_type ?? ''}`);
+  const contentTokens = stemmedMeaningfulTokensFrom(record.content);
+  const metadataTokens = stemmedMeaningfulTokensFrom(`${record.kind ?? ''} ${(record.tags ?? []).join(' ')}`);
+  for (const token of stemmedMeaningfulTokensFrom(`${record.role ?? ''} ${record.session_id ?? ''} ${record.dialog_id ?? ''} ${record.turn_id ?? ''}`)) {
+    metadataTokens.add(token);
+  }
+
+  const contentMatches = countSetIntersection(queryTokens, contentTokens);
+  const metadataMatches = countSetIntersection(queryTokens, metadataTokens);
+  const categoryMatches = countSetIntersection(categoryTokens, metadataTokens) + countSetIntersection(categoryTokens, contentTokens);
+  const roleMatches = countSetIntersection(roleHintsFrom(question), roleHintsFrom(record.role));
+  const querySessionHints = sessionHintsFrom(question);
+  const recordSessions = recordSessionTokens(record);
+  const sessionMatches = countSetIntersection(querySessionHints, recordSessions);
+  const queryDateHints = dateHintsFrom(question);
+  const temporalMatches = countSetIntersection(queryDateHints, dateHintsFrom(`${record.content} ${(record.tags ?? []).join(' ')}`));
+  const temporalRecencyScore = hasRecencyIntent(queryDateHints) ? Math.min(6, Math.log2(record.ordinal + 2)) : 0;
+  const phraseScore = phraseAndProximityScore(stemmedTokenSequenceFrom(question), stemmedTokenSequenceFrom(record.content));
+
+  return {
+    record,
+    score: (contentMatches * 12)
+      + (metadataMatches * 4)
+      + (categoryMatches * 3)
+      + (roleMatches * 18)
+      + (sessionMatches * 16)
+      + (temporalMatches * 8)
+      + temporalRecencyScore
+      + phraseScore,
+    contentMatches,
+    metadataMatches,
+    categoryMatches,
+    roleMatches,
+    sessionMatches,
+    temporalMatches,
+    temporalRecencyScore,
+    phraseScore,
+  };
+}
+
+function compareRecordId(left, right) {
+  return String(left.id).localeCompare(String(right.id));
+}
+
+function compareRecencyDesc(left, right) {
+  if (left.ordinal !== right.ordinal) return right.ordinal - left.ordinal;
+  return compareRecordId(left, right);
+}
+
+function rankedByOverlap(records, query) {
+  const queryTokens = meaningfulTokensFrom(query);
+  if (queryTokens.size === 0) return [];
+  const scored = [];
+  for (const record of records) {
+    const score = tokenOverlapScore(queryTokens, record);
+    if (score > 0) scored.push({ record, score });
+  }
+  scored.sort((left, right) => {
+    if (left.score !== right.score) return right.score - left.score;
+    return compareRecencyDesc(left.record, right.record);
+  });
+  return scored.map((item) => item.record);
+}
+
+function rankedByEnigmaRelevance(records, query) {
+  const scored = [];
+  for (const record of records) {
+    const item = enigmaRelevanceScore(query, record);
+    if (item.score > 0) scored.push(item);
+  }
+  if (scored.length === 0) return [...records].sort(compareRecencyDesc);
+  scored.sort((left, right) => {
+    if (left.score !== right.score) return right.score - left.score;
+    if (left.phraseScore !== right.phraseScore) return right.phraseScore - left.phraseScore;
+    if (left.contentMatches !== right.contentMatches) return right.contentMatches - left.contentMatches;
+    if (left.roleMatches !== right.roleMatches) return right.roleMatches - left.roleMatches;
+    if (left.sessionMatches !== right.sessionMatches) return right.sessionMatches - left.sessionMatches;
+    if (left.temporalMatches !== right.temporalMatches) return right.temporalMatches - left.temporalMatches;
+    if (left.temporalRecencyScore !== right.temporalRecencyScore) return right.temporalRecencyScore - left.temporalRecencyScore;
+    if (left.metadataMatches !== right.metadataMatches) return right.metadataMatches - left.metadataMatches;
+    return compareRecencyDesc(left.record, right.record);
+  });
+  return scored.map((item) => item.record);
+}
+
+function selectRecords(methodId, records, query, topK) {
+  if (methodId === 'full_context') return records;
+  if (methodId === 'recency_last_n') return [...records].sort(compareRecencyDesc).slice(0, topK);
+  if (methodId === 'keyword_filter') return rankedByOverlap(records, query.question).slice(0, topK);
+  if (methodId === 'enigma_relevance') return rankedByEnigmaRelevance(records, query).slice(0, topK);
+  throw new Error(`Unknown method ${methodId}`);
+}
+
+function estimatePromptTokens(question, selectedRecords) {
+  let tokens = estimateTextTokens(question);
+  for (const record of selectedRecords) tokens += record.estimated_tokens;
+  return tokens;
+}
+
+function percentile(values, ratio) {
+  if (values.length === 0) return 0;
+  const sorted = [...values].sort((left, right) => left - right);
+  const index = Math.min(sorted.length - 1, Math.max(0, Math.ceil(sorted.length * ratio) - 1));
+  return sorted[index];
+}
+
+function latencySummary(values) {
+  return {
+    samples: values.length,
+    p50_ms: round6(percentile(values, 0.5)),
+    p95_ms: round6(percentile(values, 0.95)),
+    min_ms: round6(values.length === 0 ? 0 : Math.min(...values)),
+    max_ms: round6(values.length === 0 ? 0 : Math.max(...values)),
+  };
+}
+
+function rate(numerator, denominator) {
+  if (denominator === 0) return null;
+  return round6(numerator / denominator);
+}
+
+function round6(value) {
+  return Number(value.toFixed(6));
+}
+
+function mean(total, count) {
+  return count === 0 ? 0 : round6(total / count);
+}
+
+function recordFromContent(args) {
+  const content = String(args.content ?? '');
+  return {
+    id: args.id,
+    dataset_item_id: args.dataset_item_id,
+    session_id: args.session_id,
+    turn_id: args.turn_id,
+    dialog_id: args.dialog_id,
+    role: args.role,
+    kind: args.kind,
+    tags: args.tags ?? [],
+    has_answer: args.has_answer === true,
+    ordinal: args.ordinal,
+    content,
+    estimated_tokens: estimateTextTokens(content),
+  };
+}
+
+function parseLocomoEvidenceLabels(evidence) {
+  const labels = new Set();
+  const stack = Array.isArray(evidence) ? [...evidence] : [evidence];
+  while (stack.length > 0) {
+    const value = stack.shift();
+    if (Array.isArray(value)) {
+      stack.push(...value);
+      continue;
+    }
+    if (value === undefined || value === null) continue;
+    for (const match of String(value).matchAll(/D\s*(\d+)\s*:\s*(\d+)/giu)) {
+      labels.add(`D${Number(match[1])}:${Number(match[2])}`);
+    }
+  }
+  return labels;
+}
+
+function dialogIdForTurn(sessionNumber, turn, turnIndex) {
+  const raw = turn?.dia_id ?? turn?.dialog_id ?? turn?.turn_id ?? turn?.id ?? turnIndex + 1;
+  const text = String(raw);
+  const match = text.match(/^D\s*(\d+)\s*:\s*(\d+)$/iu);
+  if (match) return `D${Number(match[1])}:${Number(match[2])}`;
+  const numeric = text.match(/\d+/u)?.[0] ?? String(turnIndex + 1);
+  return `D${sessionNumber}:${Number(numeric)}`;
+}
+
+export function parseLocomoDataset(data, options = {}) {
+  const maxQa = optionalPositiveInteger(options.max_qa ?? options.maxQa, 'max_locomo_qa');
+  const rows = normalizeDatasetArray(data, 'LoCoMo');
+  const records = [];
+  const queries = [];
+  let ordinal = 0;
+  for (let sampleIndex = 0; sampleIndex < rows.length; sampleIndex += 1) {
+    const sample = rows[sampleIndex] ?? {};
+    const itemId = String(sample.sample_id ?? sample.id ?? `sample_${sampleIndex + 1}`);
+    const conversation = sample.conversation ?? {};
+    const sessionNames = Object.keys(conversation)
+      .map((key) => {
+        const match = key.match(/^session_(\d+)$/u);
+        return match ? { key, sessionNumber: Number(match[1]) } : null;
+      })
+      .filter(Boolean)
+      .sort((left, right) => left.sessionNumber - right.sessionNumber);
+    for (const { key, sessionNumber } of sessionNames) {
+      const session = conversation[key];
+      if (!Array.isArray(session)) continue;
+      for (let turnIndex = 0; turnIndex < session.length; turnIndex += 1) {
+        const turn = session[turnIndex] ?? {};
+        const dialogId = dialogIdForTurn(sessionNumber, turn, turnIndex);
+        records.push(recordFromContent({
+          id: `locomo:${itemId}:${dialogId}`,
+          dataset_item_id: itemId,
+          session_id: `D${sessionNumber}`,
+          turn_id: dialogId,
+          dialog_id: dialogId,
+          role: turn.speaker ?? turn.role ?? undefined,
+          kind: 'locomo_conversation_turn',
+          tags: ['locomo', `session_${sessionNumber}`],
+          ordinal,
+          content: turn.text ?? turn.content ?? turn.message ?? '',
+        }));
+        ordinal += 1;
+      }
+    }
+    const qaRows = Array.isArray(sample.qa) ? sample.qa : [];
+    for (let qaIndex = 0; qaIndex < qaRows.length; qaIndex += 1) {
+      if (maxQa !== undefined && queries.length >= maxQa) break;
+      const qa = qaRows[qaIndex] ?? {};
+      const evidenceDialogIds = parseLocomoEvidenceLabels(qa.evidence);
+      queries.push({
+        id: `locomo:${itemId}:qa_${qaIndex + 1}`,
+        dataset_item_id: itemId,
+        question: String(qa.question ?? ''),
+        category: qa.category === undefined ? undefined : String(qa.category),
+        evidence_dialog_ids: evidenceDialogIds,
+        abstention: evidenceDialogIds.size === 0,
+      });
+    }
+    if (maxQa !== undefined && queries.length >= maxQa) break;
+  }
+  return {
+    id: 'locomo',
+    label: 'LoCoMo',
+    source_url: LOCOMO_SOURCE_URL,
+    license: 'CC BY-NC 4.0',
+    parser: 'conversation session turns as memory records; qa evidence labels mapped to dialog ids such as D1:3 and semicolon-separated labels',
+    task_categories: ['multi-session QA', 'event summarization', 'multimodal generation over long conversations'],
+    records,
+    queries,
+  };
+}
+
+function sessionIdAt(ids, index) {
+  if (Array.isArray(ids) && ids[index] !== undefined && ids[index] !== null) return String(ids[index]);
+  return String(index);
+}
+
+function answerSessionIds(item) {
+  if (!Array.isArray(item.answer_session_ids)) return new Set();
+  return new Set(item.answer_session_ids.map((id) => String(id)));
+}
+
+function isLongMemEvalAbstention(item, answerSessions) {
+  const id = String(item.question_id ?? item.id ?? '');
+  if (id.endsWith('_abs') || id.includes('_abs_')) return true;
+  if (String(item.question_type ?? '').toLowerCase().includes('abst')) return true;
+  return answerSessions.size === 0;
+}
+
+export function parseLongMemEvalDataset(data, options = {}) {
+  const maxItems = optionalPositiveInteger(options.max_items ?? options.maxItems, 'max_longmemeval_items');
+  const rows = normalizeDatasetArray(data, 'LongMemEval');
+  const records = [];
+  const queries = [];
+  let ordinal = 0;
+  const limit = maxItems === undefined ? rows.length : Math.min(rows.length, maxItems);
+  for (let itemIndex = 0; itemIndex < limit; itemIndex += 1) {
+    const item = rows[itemIndex] ?? {};
+    const itemId = String(item.question_id ?? item.id ?? `item_${itemIndex + 1}`);
+    const sessions = Array.isArray(item.haystack_sessions) ? item.haystack_sessions : [];
+    const sessionIds = item.haystack_session_ids;
+    const evidenceTurnIds = new Set();
+    for (let sessionIndex = 0; sessionIndex < sessions.length; sessionIndex += 1) {
+      const session = sessions[sessionIndex];
+      if (!Array.isArray(session)) continue;
+      const sessionId = sessionIdAt(sessionIds, sessionIndex);
+      for (let turnIndex = 0; turnIndex < session.length; turnIndex += 1) {
+        const turn = session[turnIndex] ?? {};
+        const turnId = `${sessionId}:${turnIndex}`;
+        if (turn.has_answer === true) evidenceTurnIds.add(turnId);
+        records.push(recordFromContent({
+          id: `longmemeval:${itemId}:${turnId}`,
+          dataset_item_id: itemId,
+          session_id: sessionId,
+          turn_id: turnId,
+          role: turn.role ?? undefined,
+          kind: 'longmemeval_haystack_turn',
+          tags: ['longmemeval', String(item.question_type ?? ''), `session_${sessionId}`],
+          has_answer: turn.has_answer === true,
+          ordinal,
+          content: turn.content ?? turn.text ?? turn.message ?? '',
+        }));
+        ordinal += 1;
+      }
+    }
+    const answerSessions = answerSessionIds(item);
+    queries.push({
+      id: `longmemeval:${itemId}`,
+      dataset_item_id: itemId,
+      question: String(item.question ?? ''),
+      question_type: item.question_type === undefined ? undefined : String(item.question_type),
+      evidence_turn_ids: evidenceTurnIds,
+      evidence_session_ids: answerSessions,
+      abstention: isLongMemEvalAbstention(item, answerSessions),
+    });
+  }
+  return {
+    id: 'longmemeval',
+    label: 'LongMemEval',
+    source_url: LONGMEMEVAL_SOURCE_URLS,
+    license: 'See Hugging Face dataset card and upstream LongMemEval repository for the selected cleaned file.',
+    parser: 'haystack_sessions turns as memory records; has_answer:true turns and answer_session_ids are used as evidence labels; _abs ids are evaluated as abstention cases',
+    task_categories: ['information extraction', 'multi-session reasoning', 'temporal reasoning', 'knowledge updates', 'abstention'],
+    records,
+    queries,
+  };
+}
+
+function scoreLocomoMethod(method, dataset, topK) {
+  const latencies = [];
+  let evidenceQuestions = 0;
+  let hits = 0;
+  let exactCoverage = 0;
+  let totalTokens = 0;
+  let selectedTotal = 0;
+  for (const query of dataset.queries) {
+    const start = performance.now();
+    const records = dataset.records.filter((record) => record.dataset_item_id === query.dataset_item_id);
+    const selected = selectRecords(method.id, records, query, topK);
+    latencies.push(performance.now() - start);
+    const selectedDialogs = new Set(selected.map((record) => record.dialog_id));
+    selectedTotal += selected.length;
+    totalTokens += estimatePromptTokens(query.question, selected);
+    if (query.evidence_dialog_ids.size > 0) {
+      evidenceQuestions += 1;
+      let covered = 0;
+      for (const evidenceId of query.evidence_dialog_ids) if (selectedDialogs.has(evidenceId)) covered += 1;
+      if (covered > 0) hits += 1;
+      if (covered === query.evidence_dialog_ids.size) exactCoverage += 1;
+    }
+  }
+  return {
+    id: method.id,
+    method: method.id,
+    local_method_only: true,
+    external_provider_called: false,
+    retrieval_proxy_only: true,
+    uses_top_k: method.uses_top_k,
+    top_k: method.uses_top_k ? topK : null,
+    question_count: dataset.queries.length,
+    evidence_question_count: evidenceQuestions,
+    evidence_hit_at_k: rate(hits, evidenceQuestions),
+    exact_evidence_coverage: rate(exactCoverage, evidenceQuestions),
+    estimated_prompt_tokens: {
+      total: totalTokens,
+      mean_per_question: mean(totalTokens, dataset.queries.length),
+      estimator: 'estimateTextTokens deterministic local estimator',
+    },
+    selected_memory_count: {
+      total: selectedTotal,
+      mean_per_question: mean(selectedTotal, dataset.queries.length),
+    },
+    latency: latencySummary(latencies),
+    public_question_text_included: false,
+    public_answer_text_included: false,
+    raw_conversation_text_included: false,
+  };
+}
+
+function scoreLongMemEvalMethod(method, dataset, topK) {
+  const latencies = [];
+  let turnEvidenceQuestions = 0;
+  let turnHits = 0;
+  let exactTurnCoverage = 0;
+  let sessionEvidenceQuestions = 0;
+  let sessionHits = 0;
+  let exactSessionCoverage = 0;
+  let abstentionQuestions = 0;
+  let abstentionCorrect = 0;
+  let totalTokens = 0;
+  let selectedTotal = 0;
+  for (const query of dataset.queries) {
+    const start = performance.now();
+    const records = dataset.records.filter((record) => record.dataset_item_id === query.dataset_item_id);
+    const selected = selectRecords(method.id, records, query, topK);
+    latencies.push(performance.now() - start);
+    const selectedTurns = new Set(selected.map((record) => record.turn_id));
+    const selectedSessions = new Set(selected.map((record) => record.session_id));
+    selectedTotal += selected.length;
+    totalTokens += estimatePromptTokens(query.question, selected);
+
+    if (query.abstention) {
+      abstentionQuestions += 1;
+      let selectedGold = false;
+      for (const turnId of query.evidence_turn_ids) if (selectedTurns.has(turnId)) selectedGold = true;
+      for (const sessionId of query.evidence_session_ids) if (selectedSessions.has(sessionId)) selectedGold = true;
+      if (!selectedGold) abstentionCorrect += 1;
+      continue;
+    }
+
+    if (query.evidence_turn_ids.size > 0) {
+      turnEvidenceQuestions += 1;
+      let covered = 0;
+      for (const turnId of query.evidence_turn_ids) if (selectedTurns.has(turnId)) covered += 1;
+      if (covered > 0) turnHits += 1;
+      if (covered === query.evidence_turn_ids.size) exactTurnCoverage += 1;
+    }
+    if (query.evidence_session_ids.size > 0) {
+      sessionEvidenceQuestions += 1;
+      let covered = 0;
+      for (const sessionId of query.evidence_session_ids) if (selectedSessions.has(sessionId)) covered += 1;
+      if (covered > 0) sessionHits += 1;
+      if (covered === query.evidence_session_ids.size) exactSessionCoverage += 1;
+    }
+  }
+  return {
+    id: method.id,
+    method: method.id,
+    local_method_only: true,
+    external_provider_called: false,
+    retrieval_proxy_only: true,
+    uses_top_k: method.uses_top_k,
+    top_k: method.uses_top_k ? topK : null,
+    item_count: dataset.queries.length,
+    turn_evidence_question_count: turnEvidenceQuestions,
+    turn_evidence_hit_at_k: rate(turnHits, turnEvidenceQuestions),
+    exact_turn_evidence_coverage: rate(exactTurnCoverage, turnEvidenceQuestions),
+    session_evidence_question_count: sessionEvidenceQuestions,
+    session_evidence_hit_at_k: rate(sessionHits, sessionEvidenceQuestions),
+    exact_session_evidence_coverage: rate(exactSessionCoverage, sessionEvidenceQuestions),
+    abstention_questions: abstentionQuestions,
+    abstention_correct: abstentionCorrect,
+    abstention_correctness: rate(abstentionCorrect, abstentionQuestions),
+    estimated_prompt_tokens: {
+      total: totalTokens,
+      mean_per_item: mean(totalTokens, dataset.queries.length),
+      estimator: 'estimateTextTokens deterministic local estimator',
+    },
+    selected_memory_count: {
+      total: selectedTotal,
+      mean_per_item: mean(selectedTotal, dataset.queries.length),
+    },
+    latency: latencySummary(latencies),
+    public_question_text_included: false,
+    public_answer_text_included: false,
+    raw_conversation_text_included: false,
+  };
+}
+
+function scoreDataset(dataset, topK) {
+  const methodRows = STANDARD_MEMORY_BENCHMARK_METHODS.map((method) => (
+    dataset.id === 'locomo' ? scoreLocomoMethod(method, dataset, topK) : scoreLongMemEvalMethod(method, dataset, topK)
+  ));
+  return {
+    id: dataset.id,
+    dataset: dataset.id,
+    label: dataset.label,
+    source_url: dataset.source_url,
+    license: dataset.license,
+    parser: dataset.parser,
+    task_categories: dataset.task_categories,
+    record_count: dataset.records.length,
+    question_count: dataset.queries.length,
+    item_count: dataset.queries.length,
+    raw_question_text_included: false,
+    raw_answer_text_included: false,
+    raw_conversation_text_included: false,
+    methods: methodRows,
+  };
+}
+
+export function runStandardMemoryBenchmarkSuite(options = {}) {
+  const topK = optionalPositiveInteger(options.top_k ?? options.topK, 'top_k') ?? 5;
+  const datasetRows = [];
+  if (options.locomoData !== undefined) {
+    datasetRows.push(scoreDataset(parseLocomoDataset(options.locomoData, { max_qa: options.max_locomo_qa ?? options.maxLocomoQa }), topK));
+  }
+  if (options.longMemEvalData !== undefined || options.longmemevalData !== undefined) {
+    datasetRows.push(scoreDataset(parseLongMemEvalDataset(options.longMemEvalData ?? options.longmemevalData, { max_items: options.max_longmemeval_items ?? options.maxLongMemEvalItems }), topK));
+  }
+  if (datasetRows.length === 0) throw new Error('At least one standard dataset is required: provide locomoData and/or longMemEvalData');
+  return buildSuiteReport(datasetRows, topK, options);
+}
+
+export async function runStandardMemoryBenchmarkSuiteFromFiles(options = {}) {
+  const datasetRows = [];
+  const topK = optionalPositiveInteger(options.top_k ?? options.topK, 'top_k') ?? 5;
+  if (options.locomo !== undefined || options.locomoPath !== undefined) {
+    const path = options.locomo ?? options.locomoPath;
+    const loaded = await readJsonWithSha256(path);
+    datasetRows.push({
+      ...scoreDataset(parseLocomoDataset(loaded.data, { max_qa: options.max_locomo_qa ?? options.maxLocomoQa }), topK),
+      local_file_name: publicFileName(path),
+      input_sha256: loaded.sha256,
+    });
+  }
+  if (options.longmemeval !== undefined || options.longmemevalPath !== undefined || options.longMemEvalPath !== undefined) {
+    const path = options.longmemeval ?? options.longmemevalPath ?? options.longMemEvalPath;
+    const maxItems = optionalPositiveInteger(options.max_longmemeval_items ?? options.maxLongMemEvalItems, 'max_longmemeval_items');
+    const loaded = await readLongMemEvalJsonWithSha256(path, maxItems);
+    datasetRows.push({
+      ...scoreDataset(parseLongMemEvalDataset(loaded.data, { max_items: maxItems }), topK),
+      local_file_name: publicFileName(path),
+      input_sha256: loaded.sha256,
+    });
+  }
+  if (datasetRows.length === 0) throw new Error('Provide --locomo <path> and/or --longmemeval <path>');
+  return buildSuiteReport(datasetRows, topK, options);
+}
+
+function buildSuiteReport(datasetRows, topK, options) {
+  return {
+    schema: STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA,
+    generated_at: options.generated_at ?? new Date().toISOString(),
+    package: {
+      name: 'enigma-memory',
+      version: '0.1.15',
+    },
+    public_safe: true,
+    top_k: topK,
+    source_urls: {
+      locomo: LOCOMO_SOURCE_URL,
+      longmemeval: LONGMEMEVAL_SOURCE_URLS,
+    },
+    license_and_boundary_notes: [
+      'LoCoMo source data is CC BY-NC 4.0; keep local dataset files and raw conversations out of public reports unless separately reviewed.',
+      'LongMemEval cleaned files are operator-supplied local JSON files from the upstream Hugging Face dataset repository.',
+      'Scores are retrieval/evidence proxy metrics over official dataset labels, not LLM-generated answer accuracy.',
+      'No provider APIs, hosted runtimes, competitor SDKs, or external accounts are called by this runner.',
+      'Rows are local deterministic methods only; no third-party competitor scores or benchmark-leadership claims are emitted.',
+    ],
+    benchmark_boundaries: {
+      official_dataset_files_required: true,
+      credentials_required: false,
+      external_provider_calls: false,
+      llm_answer_accuracy_scored: false,
+      retrieval_evidence_proxy_scored: true,
+      raw_question_text_included: false,
+      raw_answer_text_included: false,
+      raw_conversation_text_included: false,
+      provider_deletion_claim: false,
+      model_forgetting_claim: false,
+      roi_or_provider_invoice_savings_claim: false,
+      compliance_certification_claim: false,
+      benchmark_leadership_claim: false,
+    },
+    relevance_logic: {
+      token_extraction: 'keyword_filter uses basic lowercase /[a-z0-9]+(?:[-_][a-z0-9]+)*/ overlap after stopword removal; enigma_relevance additionally applies deterministic suffix stemming for ing, ed, and plural s forms',
+      production_alignment: 'public-safe local approximation of production query-aware retrieval using query/content stems, role/session/kind/tag hints, category/task hints, temporal/date hints, and phrase/proximity boosts; no private memory is emitted',
+      keyword_filter_fallback: 'empty result when no query/content token overlap exists',
+      enigma_relevance_fallback: 'falls back to all local candidates only when no enhanced relevance signal exists, then applies deterministic local ranking and --top-k',
+      provider_api_used: false,
+      llm_used: false,
+    },
+    local_methods: STANDARD_MEMORY_BENCHMARK_METHODS.map((method) => ({ ...method })),
+    datasets: datasetRows,
+    dataset_rows: datasetRows,
+  };
+}
+
+function usage() {
+  return `Usage: node scripts/run-standard-memory-benchmarks.mjs [--locomo <path>] [--longmemeval <path>] [--max-locomo-qa <n>] [--max-longmemeval-items <n>] [--top-k <n>] [--out <path>]\n\nProduces schema ${STANDARD_MEMORY_BENCHMARK_SUITE_SCHEMA}. Raw question, answer, and conversation text are never written to the report. With --longmemeval and --max-longmemeval-items, the local top-level JSON array is streamed for hashing and only the requested sample items are parsed.`;
+}
+
+async function main() {
+  const options = parseArgs();
+  if (options.help) {
+    console.log(usage());
+    return;
+  }
+  const report = await runStandardMemoryBenchmarkSuiteFromFiles(options);
+  const serialized = `${JSON.stringify(report, null, 2)}\n`;
+  if (options.out) {
+    const outPath = resolve(options.out);
+    await mkdir(dirname(outPath), { recursive: true });
+    await writeFile(outPath, serialized);
+  } else {
+    process.stdout.write(serialized);
+  }
+}
+
+const invokedPath = process.argv[1] ? resolve(process.argv[1]) : '';
+const modulePath = fileURLToPath(import.meta.url);
+if (invokedPath === modulePath) {
+  main().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exitCode = 1;
+  });
+}