claude-memory-layer 1.0.26 → 1.0.28

package/memory/tool_observation/uncategorized/2026-03-05.md

@@ -1,29 +0,0 @@
-
-## 2026-03-04T15:00:17.516Z | ad4e29ab-d52e-440a-98df-40a3ca077562
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Edit","toolInput":{"file_path":"/home/walter/workspace/opensource/claude-memory-layer/src/server/api/utils.ts","old_string":"export function getServiceFromQuery(c: Context): MemoryService {\n  const project = c.req.query('project');","new_string":"export function getServiceFromQuery(c: Context): MemoryService {\n  const project = c.req.query('project') || c.req.query('projectId');","replace_all":false},"toolOutput":"{\"filePath\":\"/home/walter/workspace/opensource/claude-memory-layer/src/server/api/utils.ts\",\"oldString\":\"export function getServiceFromQuery(c: Context): MemoryService {\\n  const project = c.req.query('project');\",\"newString\":\"export function getServiceFromQuery(c: Context): MemoryService {\\n  const project = c.req.query('project') || c.req.query('projectId');\",\"originalFile\":\"/**\\n * API Utilities\\n * Shared helpers for API endpoints\\n */\\n\\nimport type { Context } from 'hono';\\nimport * as path from 'path';\\nimport * as os from 'os';\\nimport { getReadOnlyMemoryService } from '../../services/memory-service.js';\\nimport { MemoryService } from '../../services/memory-service.js';\\n\\n/**\\n * Get the appropriate MemoryService based on the ?project= query parameter.\\n * - If ?project=<hash> is set (8 hex chars), resolves directly to project storage\\n * - If ?project=<path> is set, computes hash from path\\n * - Otherwise, returns the global read-only service\\n *\\n * Always creates read-only services for the dashboard API to avoid\\n * VectorWorker lifecycle issues with per-request services.\\n */\\nexport function getServiceFromQuery(c: Context): MemoryService {\\n  const project = c.req.query('project');\\n  if (project) {\\n    // Check if it's a hash (8 hex chars) or a path\\n    const isHash = /^[a-f0-9]{8}$/.test(project);\\n    let storagePath: string;\\n\\n    if (isHash) {\\n      storagePath = path.join(os.homedir(), '.claude-code', 'memory', 'projects', project);\\n    } else {\\n      // Import hashProjectPath dynamically to compute the hash from path\\n      const crypto = require('crypto');\\n      const normalized = project.replace(/\\\\/+$/, '') || '/';\\n      const hash = crypto.createHash('sha256').update(normalized).digest('hex').slice(0, 8);\\n      storagePath = path.join(os.homedir(), '.claude-code', 'memory', 'projects', hash);\\n    }\\n\\n    return new MemoryService({\\n      storagePath,\\n      readOnly: true,\\n      analyticsEnabled: false,\\n      sharedStoreConfig: { enabled: false }\\n    });\\n  }\\n  return getReadOnlyMemoryService();\\n}\\n\",\"structuredPatch\":[{\"oldStart\":19,\"oldLines\":7,\"newStart\":19,\"newLines\":7,\"lines\":[\"  * VectorWorker lifecycle issues with per-request services.\",\"  */\",\" export function getServiceFromQuery(c: Context): MemoryService {\",\"-  const project = c.req.query('project');\",\"+  const project = c.req.query('project') || c.req.query('projectId');\",\"   if (project) {\",\"     // Check if it's a hash (8 hex chars) or a path\",\"     const isHash = /^[a-f0-9]{8}$/.test(project);\"]}],\"userModified\":false,\"replaceAll\":false}","durationMs":0,"success":true,"metadata":{"filePath":"/home/walter/workspace/opensource/claude-memory-layer/src/server/api/utils.ts","fileType":"typescript","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:00:30.869Z | 3c248d45-27b2-48fe-af1f-448078eeb52d
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Edit","toolInput":{"file_path":"/home/walter/workspace/opensource/claude-memory-layer/src/core/sqlite-event-store.ts","old_string":"      // Calculate helpfulness score\n      const retrievalScore = retrieval.retrieval_score as number || 0;\n      const helpfulnessScore = (\n        0.30 * Math.min(retrievalScore, 1.0) +\n        0.25 * (sessionContinued ? 1.0 : 0.0) +\n        0.25 * toolSuccessRatio +\n        0.20 * (wasReasked ? 0.0 : 1.0)\n      );","new_string":"      // Calculate helpfulness score\n      // Weights tuned for shopping-assistant-like corpora where sessions\n      // continue on the same topic (was_reasked was over-penalising normal conversation flow)\n      const retrievalScore = retrieval.retrieval_score as number || 0;\n      // More prompts after retrieval = memory was actually useful to the conversation\n      const promptNorm = Math.min(promptCountAfter / 2, 1.0);\n      const helpfulnessScore = (\n        0.40 * Math.min(retrievalScore, 1.0) +\n        0.30 * promptNorm +\n        0.20 * toolSuccessRatio +\n        0.10 * (sessionContinued ? 1.0 : 0.0)\n      );","replace_all":false},"toolOutput":"{\"filePath\":\"/home/walter/workspace/opensource/claude-memory-layer/src/core/sqlite-event-store.ts\",\"oldString\":\"      // Calculate helpfulness score\\n      const retrievalScore = retrieval.retrieval_score as number || 0;\\n      const helpfulnessScore = (\\n        0.30 * Math.min(retrievalScore, 1.0) +\\n        0.25 * (sessionContinued ? 1.0 : 0.0) +\\n        0.25 * toolSuccessRatio +\\n        0.20 * (wasReasked ? 0.0 : 1.0)\\n      );\",\"newString\":\"      // Calculate helpfulness score\\n      // Weights tuned for shopping-assistant-like corpora where sessions\\n      // continue on the same topic (was_reasked was over-penalising normal conversation flow)\\n      const retrievalScore = retrieval.retrieval_score as number || 0;\\n      // More prompts after retrieval = memory was actually useful to the conversation\\n      const promptNorm = Math.min(promptCountAfter / 2, 1.0);\\n      const helpfulnessScore = (\\n        0.40 * Math.min(retrievalScore, 1.0) +\\n        0.30 * promptNorm +\\n        0.20 * toolSuccessRatio +\\n        0.10 * (sessionContinued ? 1.0 : 0.0)\\n      );\",\"originalFile\":\"/**\\n * SQLite-based EventStore implementation\\n * Primary store for hooks - WAL mode enables concurrent access\\n */\\n\\nimport { randomUUID } from 'crypto';\\nimport {\\n  MemoryEvent,\\n  MemoryEventInput,\\n  Session,\\n  AppendResult,\\n  OutboxItem\\n} from './types.js';\\nimport { makeCanonicalKey, makeDedupeKey } from './canonical-key.js';\\nimport {\\n  createSQLiteDatabase,\\n  sqliteRun,\\n  sqliteAll,\\n  sqliteGet,\\n  sqliteClose,\\n  sqliteExec,\\n  toDateFromSQLite,\\n  toSQLiteTimestamp,\\n  type SQLiteDatabase,\\n  type SQLiteOptions\\n} from './sqlite-wrapper.js';\\nimport { MarkdownMirror } from './markdown-mirror.js';\\n\\nexport interface SQLiteEventStoreOptions extends SQLiteOptions {\\n  markdownMirrorRoot?: string;\\n}\\n\\nexport class SQLiteEventStore {\\n  private db: SQLiteDatabase;\\n  private initialized = false;\\n  private readonly readOnly: boolean;\\n  private readonly markdownMirror: MarkdownMirror | null;\\n\\n  constructor(private dbPath: string, options?: SQLiteEventStoreOptions) {\\n    this.readOnly = options?.readonly ?? false;\\n    this.db = createSQLiteDatabase(dbPath, {\\n      readonly: this.readOnly,\\n      walMode: !this.readOnly\\n    });\\n    this.markdownMirror = this.readOnly || !options?.markdownMirrorRoot\\n      ? null\\n      : new MarkdownMirror(options.markdownMirrorRoot);\\n  }\\n\\n  /**\\n   * Initialize database schema\\n   */\\n  async initialize(): Promise<void> {\\n    if (this.initialized) return;\\n\\n    // In read-only mode, skip schema creation\\n    if (this.readOnly) {\\n      this.initialized = true;\\n      return;\\n    }\\n\\n    // Create all tables in a single exec for efficiency\\n    sqliteExec(this.db, `\\n      -- L0 EventStore: Single Source of Truth (immutable, append-only)\\n      CREATE TABLE IF NOT EXISTS events (\\n        id TEXT PRIMARY KEY,\\n        event_type TEXT NOT NULL,\\n        session_id TEXT NOT NULL,\\n        timestamp TEXT NOT NULL,\\n        content TEXT NOT NULL,\\n        canonical_key TEXT NOT NULL,\\n        dedupe_key TEXT UNIQUE,\\n        metadata TEXT,\\n        access_count INTEGER DEFAULT 0,\\n        last_accessed_at TEXT\\n      );\\n\\n      -- Dedup table for idempotency\\n      CREATE TABLE IF NOT EXISTS event_dedup (\\n        dedupe_key TEXT PRIMARY KEY,\\n        event_id TEXT NOT NULL,\\n        created_at TEXT DEFAULT (datetime('now'))\\n      );\\n\\n      -- Session metadata\\n      CREATE TABLE IF NOT EXISTS sessions (\\n        id TEXT PRIMARY KEY,\\n        started_at TEXT NOT NULL,\\n        ended_at TEXT,\\n        project_path TEXT,\\n        summary TEXT,\\n        tags TEXT\\n      );\\n\\n      -- Insights (derived data, rebuildable)\\n      CREATE TABLE IF NOT EXISTS insights (\\n        id TEXT PRIMARY KEY,\\n        insight_type TEXT NOT NULL,\\n        content TEXT NOT NULL,\\n        canonical_key TEXT NOT NULL,\\n        confidence REAL,\\n        source_events TEXT,\\n        created_at TEXT,\\n        last_updated TEXT\\n      );\\n\\n      -- Embedding Outbox (Single-Writer Pattern)\\n      CREATE TABLE IF NOT EXISTS embedding_outbox (\\n        id TEXT PRIMARY KEY,\\n        event_id TEXT NOT NULL,\\n        content TEXT NOT NULL,\\n        status TEXT DEFAULT 'pending',\\n        retry_count INTEGER DEFAULT 0,\\n        created_at TEXT DEFAULT (datetime('now')),\\n        processed_at TEXT,\\n        error_message TEXT\\n      );\\n\\n      -- Projection offset tracking\\n      CREATE TABLE IF NOT EXISTS projection_offsets (\\n        projection_name TEXT PRIMARY KEY,\\n        last_event_id TEXT,\\n        last_timestamp TEXT,\\n        updated_at TEXT DEFAULT (datetime('now'))\\n      );\\n\\n      -- Memory level tracking\\n      CREATE TABLE IF NOT EXISTS memory_levels (\\n        event_id TEXT PRIMARY KEY,\\n        level TEXT NOT NULL DEFAULT 'L0',\\n        promoted_at TEXT DEFAULT (datetime('now'))\\n      );\\n\\n      -- Entries (immutable memory units)\\n      CREATE TABLE IF NOT EXISTS entries (\n... [49537 characters truncated] ...\n {\\n        const metadata = JSON.parse(row.metadata);\\n        if (metadata.turnId) {\\n          sqliteRun(\\n            this.db,\\n            `UPDATE events SET turn_id = ? WHERE id = ?`,\\n            [metadata.turnId, row.id]\\n          );\\n          updated++;\\n        }\\n      } catch {\\n        // Skip rows with invalid JSON\\n      }\\n    }\\n\\n    return updated;\\n  }\\n\\n  /**\\n   * Delete all events for a session (for force reimport)\\n   */\\n  async deleteSessionEvents(sessionId: string): Promise<number> {\\n    await this.initialize();\\n\\n    // Get event IDs first for cascading deletes\\n    const events = sqliteAll<{ id: string }>(\\n      this.db,\\n      `SELECT id FROM events WHERE session_id = ?`,\\n      [sessionId]\\n    );\\n\\n    if (events.length === 0) return 0;\\n\\n    const eventIds = events.map(e => e.id);\\n    const placeholders = eventIds.map(() => '?').join(',');\\n\\n    // Drop FTS triggers to prevent SQLITE_CORRUPT_VTAB during bulk delete\\n    const ftsTriggersDropped: string[] = [];\\n    for (const triggerName of ['events_fts_delete', 'events_fts_update', 'events_fts_insert']) {\\n      try {\\n        sqliteRun(this.db, `DROP TRIGGER IF EXISTS ${triggerName}`);\\n        ftsTriggersDropped.push(triggerName);\\n      } catch {\\n        // Trigger may not exist\\n      }\\n    }\\n\\n    // Delete from related tables first (some may not exist depending on DB version)\\n    for (const table of ['event_dedup', 'memory_levels', 'embedding_queue', 'embedding_outbox', 'vector_outbox']) {\\n      try {\\n        sqliteRun(this.db, `DELETE FROM ${table} WHERE event_id IN (${placeholders})`, eventIds);\\n      } catch {\\n        // Table may not exist\\n      }\\n    }\\n\\n    // Delete events\\n    const result = sqliteRun(this.db, `DELETE FROM events WHERE session_id = ?`, [sessionId]);\\n\\n    // Rebuild FTS index if we dropped triggers\\n    if (ftsTriggersDropped.length > 0) {\\n      try {\\n        // Rebuild FTS from remaining events\\n        sqliteRun(this.db, `INSERT INTO events_fts(events_fts) VALUES('rebuild')`);\\n\\n        // Recreate triggers\\n        sqliteRun(this.db, `CREATE TRIGGER IF NOT EXISTS events_fts_insert AFTER INSERT ON events BEGIN\\n          INSERT INTO events_fts(rowid, content) VALUES (NEW.rowid, NEW.content);\\n        END`);\\n        sqliteRun(this.db, `CREATE TRIGGER IF NOT EXISTS events_fts_delete AFTER DELETE ON events BEGIN\\n          INSERT INTO events_fts(events_fts, rowid, content) VALUES('delete', OLD.rowid, OLD.content);\\n        END`);\\n        sqliteRun(this.db, `CREATE TRIGGER IF NOT EXISTS events_fts_update AFTER UPDATE ON events BEGIN\\n          INSERT INTO events_fts(events_fts, rowid, content) VALUES('delete', OLD.rowid, OLD.content);\\n          INSERT INTO events_fts(rowid, content) VALUES (NEW.rowid, NEW.content);\\n        END`);\\n      } catch {\\n        // FTS rebuild failed - non-critical, will be rebuilt on next initialize\\n      }\\n    }\\n\\n    return result.changes || 0;\\n  }\\n\\n  /**\\n   * Convert database row to MemoryEvent\\n   */\\n  private rowToEvent(row: Record<string, unknown>): MemoryEvent {\\n    const event: any = {\\n      id: row.id as string,\\n      eventType: row.event_type as 'user_prompt' | 'agent_response' | 'session_summary',\\n      sessionId: row.session_id as string,\\n      timestamp: toDateFromSQLite(row.timestamp),\\n      content: row.content as string,\\n      canonicalKey: row.canonical_key as string,\\n      dedupeKey: row.dedupe_key as string,\\n      metadata: row.metadata ? JSON.parse(row.metadata as string) : undefined\\n    };\\n\\n    // Include access tracking fields if present\\n    if (row.access_count !== undefined) {\\n      event.access_count = row.access_count;\\n    }\\n    if (row.last_accessed_at !== undefined) {\\n      event.last_accessed_at = row.last_accessed_at;\\n    }\\n    // Include turn_id if present\\n    if (row.turn_id !== undefined && row.turn_id !== null) {\\n      event.turn_id = row.turn_id;\\n    }\\n\\n    return event;\\n  }\\n}\\n\",\"structuredPatch\":[{\"oldStart\":1252,\"oldLines\":12,\"newStart\":1252,\"newLines\":16,\"lines\":[\"       }\",\" \",\"       // Calculate helpfulness score\",\"+      // Weights tuned for shopping-assistant-like corpora where sessions\",\"+      // continue on the same topic (was_reasked was over-penalising normal conversation flow)\",\"       const retrievalScore = retrieval.retrieval_score as number || 0;\",\"+      // More prompts after retrieval = memory was actually useful to the conversation\",\"+      const promptNorm = Math.min(promptCountAfter / 2, 1.0);\",\"       const helpfulnessScore = (\",\"-        0.30 * Math.min(retrievalScore, 1.0) +\",\"-        0.25 * (sessionContinued ? 1.0 : 0.0) +\",\"-        0.25 * toolSuccessRatio +\",\"-        0.20 * (wasReasked ? 0.0 : 1.0)\",\"+        0.40 * Math.min(retrievalScore, 1.0) +\",\"+        0.30 * promptNorm +\",\"+        0.20 * toolSuccessRatio +\",\"+        0.10 * (sessionContinued ? 1.0 : 0.0)\",\"       );\",\" \",\"       sqliteRun(\"]}],\"userModified\":false,\"replaceAll\":false}","durationMs":0,"success":true,"metadata":{"filePath":"/home/walter/workspace/opensource/claude-memory-layer/src/core/sqlite-event-store.ts","fileType":"typescript","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:00:40.454Z | ab79f6a8-88f2-40cf-9bf6-ffad63570206
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Edit","toolInput":{"file_path":"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts","old_string":"function isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\n  const lower = output.toLowerCase();\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\n  return output.trim().length > 800;\n}","new_string":"function isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\n  const lower = output.toLowerCase();\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\n  return output.trim().length > 2000;\n}","replace_all":false},"toolOutput":"{\"filePath\":\"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts\",\"oldString\":\"function isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\\n  const lower = output.toLowerCase();\\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\\n  return output.trim().length > 800;\\n}\",\"newString\":\"function isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\\n  const lower = output.toLowerCase();\\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\\n  return output.trim().length > 2000;\\n}\",\"originalFile\":\"#!/usr/bin/env node\\n/**\\n * PostToolUse Hook\\n * Called after each tool execution - stores tool observations\\n *\\n * Actual Claude Code input format:\\n * {\\n *   session_id, tool_name, tool_input, tool_use_id,\\n *   tool_response: { stdout?, stderr?, content?, interrupted?, isImage? },\\n *   cwd, transcript_path, permission_mode, hook_event_name\\n * }\\n */\\n\\nimport { getLightweightMemoryService } from '../services/memory-service.js';\\nimport { applyPrivacyFilter, maskSensitiveInput, truncateOutput } from '../core/privacy/index.js';\\nimport { extractMetadata } from '../core/metadata-extractor.js';\\nimport { readTurnState } from '../core/turn-state.js';\\nimport type { PostToolUseInput, ToolObservationPayload, Config } from '../core/types.js';\\n\\n// Default config\\nconst DEFAULT_CONFIG: Config['toolObservation'] = {\\n  enabled: true,\\n  excludedTools: [\\n    // Trivial meta tools\\n    'TodoWrite', 'TodoRead',\\n    // Reproducible query tools (no storage value)\\n    'Read', 'Grep', 'Glob',\\n    'ToolSearch', 'WebFetch', 'WebSearch', 'NotebookRead',\\n    // Low-value system tools\\n    'Skill', 'EnterPlanMode',\\n  ],\\n  minOutputLength: parseInt(process.env.CLAUDE_MEMORY_TOOL_MIN_OUTPUT_LEN || '100'),\\n  maxOutputLength: 10000,\\n  maxOutputLines: 100,\\n  storeOnlyOnSuccess: false\\n};\\n\\n// Tools that are always stored regardless of output length\\nconst ALWAYS_STORE_TOOLS = new Set([\\n  'Write', 'Edit', 'MultiEdit', 'Agent', 'Task', 'ExitPlanMode'\\n]);\\n\\n// Keywords that indicate a Bash output is worth storing\\nconst IMPORTANT_BASH_KEYWORDS = [\\n  'error', 'failed', 'exception', 'traceback', 'panic',\\n  'warning', 'deprecated',\\n  'test passed', 'test failed', 'tests passed', 'tests failed',\\n  'coverage', 'assert',\\n  'published', 'deployed', 'built successfully', 'build complete',\\n  'successfully installed', 'successfully created',\\n];\\n\\n/**\\n * For Bash commands, only store output that is significant:\\n * - Has stderr content\\n * - Contains important keywords (errors, test results, deploy events)\\n * - Output is very long (> 800 chars), indicating meaningful work\\n */\\nfunction isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\\n  const lower = output.toLowerCase();\\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\\n  return output.trim().length > 800;\\n}\\n\\n/**\\n * Determine if a tool output is significant enough to store.\\n * Always-store tools bypass the length check.\\n * Bash uses keyword-based significance detection.\\n * Other tools require non-empty stderr or output length >= minLen.\\n */\\nfunction hasSignificantOutput(\\n  toolName: string,\\n  output: string,\\n  response: PostToolUseInput['tool_response'],\\n  minLen: number\\n): boolean {\\n  if (ALWAYS_STORE_TOOLS.has(toolName)) return true;\\n  if (toolName === 'Bash') return isBashSignificant(output, response);\\n  if (response?.stderr && response.stderr.trim().length > 0) return true;\\n  return output.trim().length >= minLen;\\n}\\n\\nconst DEFAULT_PRIVACY_CONFIG: Config['privacy'] = {\\n  excludePatterns: ['password', 'secret', 'api_key', 'token', 'bearer'],\\n  anonymize: false,\\n  privateTags: {\\n    enabled: true,\\n    marker: '[PRIVATE]\n',\\n    preserveLineCount: false,\\n    supportedFormats: ['xml']\\n  }\\n};\\n\\n/**\\n * Extract text output from tool_response object\\n */\\nfunction extractToolOutput(response: PostToolUseInput['tool_response']): string {\\n  if (!response) return '';\\n\\n  // Bash tools: stdout + stderr\\n  if (response.stdout !== undefined) {\\n    const parts: string[] = [];\\n    if (response.stdout) parts.push(response.stdout);\\n    if (response.stderr) parts.push(`[stderr] ${response.stderr}`);\\n    return parts.join('\\\\n') || '';\\n  }\\n\\n  // Other tools may have content field\\n  if (response.content !== undefined) {\\n    return typeof response.content === 'string'\\n      ? response.content\\n      : JSON.stringify(response.content);\\n  }\\n\\n  // Fallback: stringify the whole response\\n  return JSON.stringify(response);\\n}\\n\\n/**\\n * Determine if the tool execution was successful\\n */\\nfunction isToolSuccess(response: PostToolUseInput['tool_response']): boolean {\\n  if (!response) return false;\\n  if (response.interrupted) return false;\\n  // If stderr has content but stdout also has content, still consider success\\n  return true;\\n}\\n\\nasync function main(): Promise<void> {\\n  // Read input from stdin\\n  const inputData = await readStdin();\\n  const input: PostToolUseInput = JSON.parse(inputData);\\n\\n  const config = { ...DEFAULT_CONFIG };\\n  const privacyConfig = DEFAULT_PRIVACY_CONFIG;\\n\\n  // Allow env-based blocklist override\\n  const envBlocklist = process.env.CLAUDE_MEMORY_TOOL_BLOCKLIST;\\n  if (envBlocklist !== undefined) {\\n    config.excludedTools = envBlocklist.split(',').map((s) => s.trim()).filter(Boolean);\\n  }\\n\\n  // 1. Check if tool observation is enabled\\n  if (!config.enabled) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 2. Check if tool is excluded\\n  if (config.excludedTools?.includes(input.tool_name)) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 3. Extract output from tool_response object\\n  const toolOutput = extractToolOutput(input.tool_response);\\n  const success = isToolSuccess(input.tool_response);\\n\\n  // 4. Check success filter\\n  if (!success && config.storeOnlyOnSuccess) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 4.5. Output-level filter: skip low-signal outputs\\n  if (!hasSignificantOutput(\\n    input.tool_name, toolOutput, input.tool_response,\\n    config.minOutputLength ?? 100\\n  )) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  try {\\n    const memoryService = getLightweightMemoryService(input.session_id);\\n\\n    // 5. Mask sensitive data in input\\n    const maskedInput = maskSensitiveInput(input.tool_input);\\n\\n    // 6. Apply privacy filter to output\\n    const filterResult = applyPrivacyFilter(toolOutput, privacyConfig);\\n    const maskedOutput = filterResult.content;\\n\\n    // 7. Truncate output\\n    const truncatedOutput = truncateOutput(maskedOutput, {\\n      maxLength: config.maxOutputLength,\\n      maxLines: config.maxOutputLines\\n    });\\n\\n    // 8. Extract metadata\\n    const metadata = extractMetadata(\\n      input.tool_name,\\n      maskedInput,\\n      toolOutput,\\n      success\\n    );\\n\\n    // 8.5. Read current turn_id from state file\\n    const turnId = readTurnState(input.session_id);\\n\\n    // 9. Create payload (include turnId in metadata for grouping)\\n    const payload: ToolObservationPayload = {\\n      toolName: input.tool_name,\\n      toolInput: maskedInput,\\n      toolOutput: truncatedOutput,\\n      durationMs: 0, // Claude Code doesn't provide timing info\\n      success,\\n      errorMessage: input.tool_response?.stderr || undefined,\\n      metadata: {\\n        ...metadata,\\n        ...(turnId ? { turnId } : {})\\n      }\\n    };\\n\\n    // 10. Store observation\\n    await memoryService.storeToolObservation(input.session_id, payload);\\n\\n    // Output empty (hook doesn't return context)\\n    console.log(JSON.stringify({}));\\n  } catch (error) {\\n    if (process.env.CLAUDE_MEMORY_DEBUG) {\\n      console.error('PostToolUse hook error:', error);\\n    }\\n    console.log(JSON.stringify({}));\\n  }\\n}\\n\\nfunction readStdin(): Promise<string> {\\n  return new Promise((resolve) => {\\n    let data = '';\\n    process.stdin.setEncoding('utf8');\\n    process.stdin.on('data', (chunk) => {\\n      data += chunk;\\n    });\\n    process.stdin.on('end', () => {\\n      resolve(data);\\n    });\\n  });\\n}\\n\\nmain().catch(console.error);\\n\",\"structuredPatch\":[{\"oldStart\":60,\"oldLines\":7,\"newStart\":60,\"newLines\":7,\"lines\":[\"   if (response?.stderr && response.stderr.trim().length > 20) return true;\",\"   const lower = output.toLowerCase();\",\"   if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\",\"-  return output.trim().length > 800;\",\"+  return output.trim().length > 2000;\",\" }\",\" \",\" /**\"]}],\"userModified\":false,\"replaceAll\":false}","durationMs":0,"success":true,"metadata":{"filePath":"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts","fileType":"typescript","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:00:48.809Z | 588474cb-d13f-435b-bf4f-dd87751d9318
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Edit","toolInput":{"file_path":"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts","old_string":" * - Output is very long (> 800 chars), indicating meaningful work","new_string":" * - Output is very long (> 2000 chars), indicating meaningful work","replace_all":false},"toolOutput":"{\"filePath\":\"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts\",\"oldString\":\" * - Output is very long (> 800 chars), indicating meaningful work\",\"newString\":\" * - Output is very long (> 2000 chars), indicating meaningful work\",\"originalFile\":\"#!/usr/bin/env node\\n/**\\n * PostToolUse Hook\\n * Called after each tool execution - stores tool observations\\n *\\n * Actual Claude Code input format:\\n * {\\n *   session_id, tool_name, tool_input, tool_use_id,\\n *   tool_response: { stdout?, stderr?, content?, interrupted?, isImage? },\\n *   cwd, transcript_path, permission_mode, hook_event_name\\n * }\\n */\\n\\nimport { getLightweightMemoryService } from '../services/memory-service.js';\\nimport { applyPrivacyFilter, maskSensitiveInput, truncateOutput } from '../core/privacy/index.js';\\nimport { extractMetadata } from '../core/metadata-extractor.js';\\nimport { readTurnState } from '../core/turn-state.js';\\nimport type { PostToolUseInput, ToolObservationPayload, Config } from '../core/types.js';\\n\\n// Default config\\nconst DEFAULT_CONFIG: Config['toolObservation'] = {\\n  enabled: true,\\n  excludedTools: [\\n    // Trivial meta tools\\n    'TodoWrite', 'TodoRead',\\n    // Reproducible query tools (no storage value)\\n    'Read', 'Grep', 'Glob',\\n    'ToolSearch', 'WebFetch', 'WebSearch', 'NotebookRead',\\n    // Low-value system tools\\n    'Skill', 'EnterPlanMode',\\n  ],\\n  minOutputLength: parseInt(process.env.CLAUDE_MEMORY_TOOL_MIN_OUTPUT_LEN || '100'),\\n  maxOutputLength: 10000,\\n  maxOutputLines: 100,\\n  storeOnlyOnSuccess: false\\n};\\n\\n// Tools that are always stored regardless of output length\\nconst ALWAYS_STORE_TOOLS = new Set([\\n  'Write', 'Edit', 'MultiEdit', 'Agent', 'Task', 'ExitPlanMode'\\n]);\\n\\n// Keywords that indicate a Bash output is worth storing\\nconst IMPORTANT_BASH_KEYWORDS = [\\n  'error', 'failed', 'exception', 'traceback', 'panic',\\n  'warning', 'deprecated',\\n  'test passed', 'test failed', 'tests passed', 'tests failed',\\n  'coverage', 'assert',\\n  'published', 'deployed', 'built successfully', 'build complete',\\n  'successfully installed', 'successfully created',\\n];\\n\\n/**\\n * For Bash commands, only store output that is significant:\\n * - Has stderr content\\n * - Contains important keywords (errors, test results, deploy events)\\n * - Output is very long (> 800 chars), indicating meaningful work\\n */\\nfunction isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\\n  if (response?.stderr && response.stderr.trim().length > 20) return true;\\n  const lower = output.toLowerCase();\\n  if (IMPORTANT_BASH_KEYWORDS.some((kw) => lower.includes(kw))) return true;\\n  return output.trim().length > 2000;\\n}\\n\\n/**\\n * Determine if a tool output is significant enough to store.\\n * Always-store tools bypass the length check.\\n * Bash uses keyword-based significance detection.\\n * Other tools require non-empty stderr or output length >= minLen.\\n */\\nfunction hasSignificantOutput(\\n  toolName: string,\\n  output: string,\\n  response: PostToolUseInput['tool_response'],\\n  minLen: number\\n): boolean {\\n  if (ALWAYS_STORE_TOOLS.has(toolName)) return true;\\n  if (toolName === 'Bash') return isBashSignificant(output, response);\\n  if (response?.stderr && response.stderr.trim().length > 0) return true;\\n  return output.trim().length >= minLen;\\n}\\n\\nconst DEFAULT_PRIVACY_CONFIG: Config['privacy'] = {\\n  excludePatterns: ['password', 'secret', 'api_key', 'token', 'bearer'],\\n  anonymize: false,\\n  privateTags: {\\n    enabled: true,\\n    marker: '[PRIVATE]\n',\\n    preserveLineCount: false,\\n    supportedFormats: ['xml']\\n  }\\n};\\n\\n/**\\n * Extract text output from tool_response object\\n */\\nfunction extractToolOutput(response: PostToolUseInput['tool_response']): string {\\n  if (!response) return '';\\n\\n  // Bash tools: stdout + stderr\\n  if (response.stdout !== undefined) {\\n    const parts: string[] = [];\\n    if (response.stdout) parts.push(response.stdout);\\n    if (response.stderr) parts.push(`[stderr] ${response.stderr}`);\\n    return parts.join('\\\\n') || '';\\n  }\\n\\n  // Other tools may have content field\\n  if (response.content !== undefined) {\\n    return typeof response.content === 'string'\\n      ? response.content\\n      : JSON.stringify(response.content);\\n  }\\n\\n  // Fallback: stringify the whole response\\n  return JSON.stringify(response);\\n}\\n\\n/**\\n * Determine if the tool execution was successful\\n */\\nfunction isToolSuccess(response: PostToolUseInput['tool_response']): boolean {\\n  if (!response) return false;\\n  if (response.interrupted) return false;\\n  // If stderr has content but stdout also has content, still consider success\\n  return true;\\n}\\n\\nasync function main(): Promise<void> {\\n  // Read input from stdin\\n  const inputData = await readStdin();\\n  const input: PostToolUseInput = JSON.parse(inputData);\\n\\n  const config = { ...DEFAULT_CONFIG };\\n  const privacyConfig = DEFAULT_PRIVACY_CONFIG;\\n\\n  // Allow env-based blocklist override\\n  const envBlocklist = process.env.CLAUDE_MEMORY_TOOL_BLOCKLIST;\\n  if (envBlocklist !== undefined) {\\n    config.excludedTools = envBlocklist.split(',').map((s) => s.trim()).filter(Boolean);\\n  }\\n\\n  // 1. Check if tool observation is enabled\\n  if (!config.enabled) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 2. Check if tool is excluded\\n  if (config.excludedTools?.includes(input.tool_name)) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 3. Extract output from tool_response object\\n  const toolOutput = extractToolOutput(input.tool_response);\\n  const success = isToolSuccess(input.tool_response);\\n\\n  // 4. Check success filter\\n  if (!success && config.storeOnlyOnSuccess) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  // 4.5. Output-level filter: skip low-signal outputs\\n  if (!hasSignificantOutput(\\n    input.tool_name, toolOutput, input.tool_response,\\n    config.minOutputLength ?? 100\\n  )) {\\n    console.log(JSON.stringify({}));\\n    return;\\n  }\\n\\n  try {\\n    const memoryService = getLightweightMemoryService(input.session_id);\\n\\n    // 5. Mask sensitive data in input\\n    const maskedInput = maskSensitiveInput(input.tool_input);\\n\\n    // 6. Apply privacy filter to output\\n    const filterResult = applyPrivacyFilter(toolOutput, privacyConfig);\\n    const maskedOutput = filterResult.content;\\n\\n    // 7. Truncate output\\n    const truncatedOutput = truncateOutput(maskedOutput, {\\n      maxLength: config.maxOutputLength,\\n      maxLines: config.maxOutputLines\\n    });\\n\\n    // 8. Extract metadata\\n    const metadata = extractMetadata(\\n      input.tool_name,\\n      maskedInput,\\n      toolOutput,\\n      success\\n    );\\n\\n    // 8.5. Read current turn_id from state file\\n    const turnId = readTurnState(input.session_id);\\n\\n    // 9. Create payload (include turnId in metadata for grouping)\\n    const payload: ToolObservationPayload = {\\n      toolName: input.tool_name,\\n      toolInput: maskedInput,\\n      toolOutput: truncatedOutput,\\n      durationMs: 0, // Claude Code doesn't provide timing info\\n      success,\\n      errorMessage: input.tool_response?.stderr || undefined,\\n      metadata: {\\n        ...metadata,\\n        ...(turnId ? { turnId } : {})\\n      }\\n    };\\n\\n    // 10. Store observation\\n    await memoryService.storeToolObservation(input.session_id, payload);\\n\\n    // Output empty (hook doesn't return context)\\n    console.log(JSON.stringify({}));\\n  } catch (error) {\\n    if (process.env.CLAUDE_MEMORY_DEBUG) {\\n      console.error('PostToolUse hook error:', error);\\n    }\\n    console.log(JSON.stringify({}));\\n  }\\n}\\n\\nfunction readStdin(): Promise<string> {\\n  return new Promise((resolve) => {\\n    let data = '';\\n    process.stdin.setEncoding('utf8');\\n    process.stdin.on('data', (chunk) => {\\n      data += chunk;\\n    });\\n    process.stdin.on('end', () => {\\n      resolve(data);\\n    });\\n  });\\n}\\n\\nmain().catch(console.error);\\n\",\"structuredPatch\":[{\"oldStart\":54,\"oldLines\":7,\"newStart\":54,\"newLines\":7,\"lines\":[\"  * For Bash commands, only store output that is significant:\",\"  * - Has stderr content\",\"  * - Contains important keywords (errors, test results, deploy events)\",\"- * - Output is very long (> 800 chars), indicating meaningful work\",\"+ * - Output is very long (> 2000 chars), indicating meaningful work\",\"  */\",\" function isBashSignificant(output: string, response: PostToolUseInput['tool_response']): boolean {\",\"   if (response?.stderr && response.stderr.trim().length > 20) return true;\"]}],\"userModified\":false,\"replaceAll\":false}","durationMs":0,"success":true,"metadata":{"filePath":"/home/walter/workspace/opensource/claude-memory-layer/src/hooks/post-tool-use.ts","fileType":"typescript","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:00:53.922Z | 628373c5-d886-4d2f-824b-060b2655ba04
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Bash","toolInput":{"command":"npm run build 2>&1 | tail -15","timeout":60000,"description":"Build to verify all changes compile"},"toolOutput":"📦 Building services...\n📦 Building server...\n📋 Copying plugin files...\n📋 Copying UI files...\n\n✅ Build complete!\n\nOutput: dist/\n  - cli/index.js\n  - hooks/*.js\n  - core/index.js\n  - services/memory-service.js\n  - server/index.js\n  - ui/index.html\n  - .claude-plugin/","durationMs":0,"success":true,"metadata":{"command":"npm run","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:01:15.931Z | bda6bba3-917f-4a27-9361-3f7057fcb59c
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"Edit","toolInput":{"file_path":"/home/walter/workspace/opensource/claude-memory-layer/specs/memory-utilization-improvements/spec.md","old_string":"## 구현 순서\n\n```\nWeek 1 (Critical Path):\n  IMP-01: Trace 동기 기록 → 즉시 측정 가능\n  IMP-02: Tool Observation 필터링 → 노이즈 감소\n\nWeek 2 (Quality):\n  IMP-03: 세션 요약 신뢰성\n  IMP-04: Graduation repair CLI\n  IMP-05: Embedding 모델 모니터링\n\nWeek 3 (Feedback Loop):\n  IMP-06: Helpfulness 피드백\n  IMP-07: 컨텍스트 포맷 개선\n```","new_string":"## 2차 개선 (실측 데이터 기반 후속)\n\n### IMP-01b: 대시보드 API projectId 파라미터 불일치 수정\n\n**문제**: 대시보드가 `?projectId=f4d5c120`으로 쿼리하는데 `getServiceFromQuery()`는 `?project=`만 읽음 → 항상 글로벌 서비스 반환\n\n**수정**: `src/server/api/utils.ts`\n```typescript\nconst project = c.req.query('project') || c.req.query('projectId');\n```\n**완료** ✅\n\n---\n\n### IMP-02b: Bash 출력 임계값 800 → 2000\n\n**문제**: 800자 임계값이 너무 낮아 일반적인 bash 출력 대부분이 저장됨\n\n**수정**: `src/hooks/post-tool-use.ts`\n```typescript\nreturn output.trim().length > 2000;\n```\n**완료** ✅\n\n---\n\n### IMP-06b: Helpfulness 평가 알고리즘 변별력 개선\n\n**문제**: `was_reasked` 로직이 자연스러운 대화 연속을 penalty로 처리 → 모든 항목 0.62~0.68 수렴\n\n**원인**:\n- 쇼핑 어시스턴트처럼 같은 주제를 이어가면 `was_reasked=1` 항상 발생\n- 가중치: `0.20 × (wasReasked ? 0 : 1)` → 모든 항목에 -0.20 penalty\n\n**수정**: `src/core/sqlite-event-store.ts`\n```typescript\n// 기존 (변별력 낮음)\n0.30 × retrievalScore + 0.25 × sessionContinued + 0.25 × toolSuccessRatio + 0.20 × !wasReasked\n\n// 개선 (prompt_count_after 활용)\n0.40 × retrievalScore + 0.30 × promptNorm + 0.20 × toolSuccessRatio + 0.10 × sessionContinued\n// promptNorm = min(promptCountAfter / 2, 1.0)\n// 0턴→0.0, 1턴→0.5, 2턴이상→1.0\n```\n**완료** ✅\n\n---\n\n## 구현 순서\n\n```\nWeek 1 (Critical Path) - 완료:\n  IMP-01: Trace projectHash 수정 ✅\n  IMP-02: Tool Observation 필터링 ✅\n\nWeek 2 (Quality) - 완료:\n  IMP-03: 세션 요약 신뢰성 ✅\n  IMP-04: Graduation repair CLI\n  IMP-05: Embedding 모델 모니터링\n\nWeek 2 (후속 실측 기반) - 완료:\n  IMP-01b: 대시보드 projectId 파라미터 수정 ✅\n  IMP-02b: Bash 임계값 800→2000 ✅\n  IMP-06b: Helpfulness 알고리즘 개선 ✅\n\nWeek 3 (Feedback Loop):\n  IMP-06: Helpfulness 피드백 루프 완성\n  IMP-07: 컨텍스트 포맷 개선\n```","replace_all":false},"toolOutput":"{\"filePath\":\"/home/walter/workspace/opensource/claude-memory-layer/specs/memory-utilization-improvements/spec.md\",\"oldString\":\"## 구현 순서\\n\\n```\\nWeek 1 (Critical Path):\\n  IMP-01: Trace 동기 기록 → 즉시 측정 가능\\n  IMP-02: Tool Observation 필터링 → 노이즈 감소\\n\\nWeek 2 (Quality):\\n  IMP-03: 세션 요약 신뢰성\\n  IMP-04: Graduation repair CLI\\n  IMP-05: Embedding 모델 모니터링\\n\\nWeek 3 (Feedback Loop):\\n  IMP-06: Helpfulness 피드백\\n  IMP-07: 컨텍스트 포맷 개선\\n```\",\"newString\":\"## 2차 개선 (실측 데이터 기반 후속)\\n\\n### IMP-01b: 대시보드 API projectId 파라미터 불일치 수정\\n\\n**문제**: 대시보드가 `?projectId=f4d5c120`으로 쿼리하는데 `getServiceFromQuery()`는 `?project=`만 읽음 → 항상 글로벌 서비스 반환\\n\\n**수정**: `src/server/api/utils.ts`\\n```typescript\\nconst project = c.req.query('project') || c.req.query('projectId');\\n```\\n**완료** ✅\\n\\n---\\n\\n### IMP-02b: Bash 출력 임계값 800 → 2000\\n\\n**문제**: 800자 임계값이 너무 낮아 일반적인 bash 출력 대부분이 저장됨\\n\\n**수정**: `src/hooks/post-tool-use.ts`\\n```typescript\\nreturn output.trim().length > 2000;\\n```\\n**완료** ✅\\n\\n---\\n\\n### IMP-06b: Helpfulness 평가 알고리즘 변별력 개선\\n\\n**문제**: `was_reasked` 로직이 자연스러운 대화 연속을 penalty로 처리 → 모든 항목 0.62~0.68 수렴\\n\\n**원인**:\\n- 쇼핑 어시스턴트처럼 같은 주제를 이어가면 `was_reasked=1` 항상 발생\\n- 가중치: `0.20 × (wasReasked ? 0 : 1)` → 모든 항목에 -0.20 penalty\\n\\n**수정**: `src/core/sqlite-event-store.ts`\\n```typescript\\n// 기존 (변별력 낮음)\\n0.30 × retrievalScore + 0.25 × sessionContinued + 0.25 × toolSuccessRatio + 0.20 × !wasReasked\\n\\n// 개선 (prompt_count_after 활용)\\n0.40 × retrievalScore + 0.30 × promptNorm + 0.20 × toolSuccessRatio + 0.10 × sessionContinued\\n// promptNorm = min(promptCountAfter / 2, 1.0)\\n// 0턴→0.0, 1턴→0.5, 2턴이상→1.0\\n```\\n**완료** ✅\\n\\n---\\n\\n## 구현 순서\\n\\n```\\nWeek 1 (Critical Path) - 완료:\\n  IMP-01: Trace projectHash 수정 ✅\\n  IMP-02: Tool Observation 필터링 ✅\\n\\nWeek 2 (Quality) - 완료:\\n  IMP-03: 세션 요약 신뢰성 ✅\\n  IMP-04: Graduation repair CLI\\n  IMP-05: Embedding 모델 모니터링\\n\\nWeek 2 (후속 실측 기반) - 완료:\\n  IMP-01b: 대시보드 projectId 파라미터 수정 ✅\\n  IMP-02b: Bash 임계값 800→2000 ✅\\n  IMP-06b: Helpfulness 알고리즘 개선 ✅\\n\\nWeek 3 (Feedback Loop):\\n  IMP-06: Helpfulness 피드백 루프 완성\\n  IMP-07: 컨텍스트 포맷 개선\\n```\",\"originalFile\":\"# Spec: Memory Utilization Improvements\\n\\n## 목표\\n\\nf4d5c120 프로젝트 분석을 통해 확인된 메모리 시스템의 cascade failure를 해소하고,\\n저장된 메모리가 실제로 Claude 응답에 기여하는 비율을 높인다.\\n\\n**성공 기준**:\\n- Retrieval trace 기록률 > 95% (현재 ~0%)\\n- 메모리 Graduation L1+ 비율 > 10% (현재 0%)\\n- Tool Observation 중 저장 비율 < 30% (현재 100%)\\n- 세션 요약 생성률 > 80% (현재 5%)\\n\\n---\\n\\n## 개선 항목 (우선순위 순)\\n\\n---\\n\\n### IMP-01: Retrieval Trace 동기 기록\\n\\n**우선순위**: P0 (Blocker)\\n\\n**문제**:\\n- `adherence: true` 이벤트가 존재 → 검색은 실행됨\\n- 하지만 retrieval_traces 테이블에 단 1건만 기록\\n- 비동기 trace INSERT가 hook 프로세스 종료 전에 완료되지 않는 것으로 추정\\n\\n**원인 분석**:\\n```typescript\\n// user-prompt-submit.ts 추정 흐름\\nasync function hook() {\\n  const memories = await retrieveMemories(query);   // 검색 실행\\n  await writeToStdout(memories);                    // 결과 출력 (hook 종료)\\n  await recordTrace(memories);                      // ← 여기서 프로세스가 이미 종료됨\\n}\\n```\\n\\n**해결 방안**:\\n1. `recordTrace()`를 `retrieveMemories()` 직후, stdout 출력 전에 동기적으로 실행\\n2. better-sqlite3의 동기 API 사용 (이미 사용 중) → trace INSERT를 sync로 변경\\n3. trace 기록 실패 시 stderr에 경고 출력\\n\\n**변경 파일**:\\n- `src/hooks/user-prompt-submit.ts` — trace 기록 위치를 stdout 출력 전으로 이동\\n- `src/core/sqlite-event-store.ts` — `recordRetrievalTrace()`를 동기 메서드로 변경\\n\\n**검증**:\\n```sql\\n-- 개선 후: 세션당 retrieval_traces 수 확인\\nSELECT session_id, COUNT(*) as trace_count\\nFROM retrieval_traces\\nGROUP BY session_id\\nORDER BY trace_count DESC;\\n```\\n\\n---\\n\\n### IMP-02: Tool Observation 선택적 저장\\n\\n**우선순위**: P0\\n\\n**문제**:\\n- 저장 이벤트의 93%가 tool observation\\n- 대부분 다음 세션에서 재사용 가치가 없는 ephemeral 데이터\\n- FTS5 검색 코퍼스 비대화 → 검색 노이즈 증가\\n\\n**저장 가치 기준**:\\n\\n| Tool | 저장 조건 | 이유 |\\n|------|---------|------|\\n| Bash | 오류(exit_code≠0), 또는 중요 출력 키워드 포함 | 성공적인 단순 명령은 가치 낮음 |\\n| Write | 항상 저장 | 파일 생성/수정은 중요 컨텍스트 |\\n| Edit | 항상 저장 | 코드 변경은 추적 필요 |\\n| Read | 저장 안 함 (기본값) | 파일 내용은 재현 가능 |\\n| Glob | 저장 안 함 | 디렉토리 구조는 변하지 않음 |\\n| Grep | 결과 10줄 초과 시만 저장 | 대용량 검색 결과만 의미있음 |\\n| TodoRead/Write | 저장 안 함 | 이미 제외됨 |\\n\\n**중요 출력 키워드** (Bash 저장 트리거):\\n```\\nerror, Error, ERROR, failed, Failed, FAILED,\\nwarning, Warning, exception, Exception,\\ntest passed, test failed, coverage,\\nsuccessfully deployed, build complete\\n```\\n\\n**변경 파일**:\\n- `src/hooks/post-tool-use.ts` — 툴별 저장 조건 필터링 로직 추가\\n- `src/core/metadata-extractor.ts` — 중요도 점수 계산 함수 추가\\n\\n**예상 효과**:\\n- 저장 이벤트 수 60~70% 감소\\n- 검색 정밀도 향상\\n- DB/벡터 스토어 용량 절감\\n\\n---\\n\\n### IMP-03: 세션 요약 신뢰성 개선\\n\\n**우선순위**: P1\\n\\n**문제**:\\n- 40 세션 중 2건만 요약 (5%)\\n- Stop hook은 프로세스 강제 종료 시 실행되지 않음\\n- Session-start의 백필 로직이 요약을 생성하지 않고 있음\\n\\n**해결 방안**:\\n\\n**A. Session-start 백필에서 요약 생성 추가**:\\n```typescript\\n// session-start.ts\\nasync function backfillPreviousSession(prevSessionId: string) {\\n  const events = await getSessionEvents(prevSessionId);\\n  if (events.length > 3 && !hasSummary(prevSessionId)) {\\n    const summary = await generateSummary(events);\\n    await storeSummaryEvent(prevSessionId, summary);\\n  }\\n}\\n```\\n\\n**B. LLM 없이 규칙 기반 요약 생성** (빠른 실행):\\n```\\n요약 템플릿:\\n\\\"[날짜] [N]턴 세션. 주요 작업: [user_prompt 첫\n... [977 characters truncated] ...\n니터링\\n\\n**우선순위**: P1\\n\\n**문제**:\\n- \\\"Unknown model class 'eurobert'\\\" 오류 반복\\n- \\\"RotaryEmbedding node execution errors\\\" 반복\\n- 사용자에게 오류가 노출되지만 자동 대처 없음\\n\\n**해결 방안**:\\n\\n**A. 모델 헬스체크 강화**:\\n```typescript\\n// semantic-daemon.ts\\nasync function checkModelHealth(): Promise<boolean> {\\n  try {\\n    await embed(\\\"test query\\\");\\n    return true;\\n  } catch (e) {\\n    logger.warn('Primary model failed, switching to fallback');\\n    await switchToFallbackModel();\\n    return false;\\n  }\\n}\\n```\\n\\n**B. 대시보드에 모델 상태 표시**:\\n- 현재 사용 중인 모델명\\n- 최근 임베딩 성공률\\n- 오류 로그 최근 5건\\n\\n**C. 임베딩 오류 시 키워드 전용 모드 자동 전환**:\\n- `CLAUDE_MEMORY_RETRIEVAL_MODE=keyword` 임시 설정\\n- 사용자에게 `<system>` 메시지로 알림\\n\\n**변경 파일**:\\n- `src/hooks/semantic-daemon.ts` — 헬스체크 및 자동 폴백 강화\\n- `src/server/api/stats.ts` — 모델 상태 지표 추가\\n\\n---\\n\\n### IMP-06: Helpfulness 피드백 루프\\n\\n**우선순위**: P2\\n\\n**문제**:\\n- `sessions_helpfulness` 테이블 존재하나 데이터 없음\\n- 검색된 메모리가 실제로 유용했는지 측정 불가\\n- 유용하지 않은 메모리가 계속 상위에 노출될 수 있음\\n\\n**해결 방안**:\\n\\n**A. 휴리스틱 기반 자동 helpfulness 평가**:\\n```\\nStop hook에서:\\n1. 이번 세션에서 검색된 eventId 목록 로드\\n2. agent_response에서 검색된 메모리의 canonicalKey가 언급되었는지 확인\\n3. 언급된 경우 helpfulness = 0.8, 아닌 경우 0.2로 기록\\n```\\n\\n**B. Retrieval Score 가중치에 helpfulness 반영**:\\n```typescript\\n// matcher.ts\\nscore = 0.35 × semanticScore +\\n        0.25 × ftsScore +\\n        0.20 × recencyBonus +\\n        0.10 × statusMultiplier +\\n        0.10 × helpfulnessScore;  // ← 신규 추가\\n```\\n\\n**C. 낮은 helpfulness 이벤트 강등**:\\n- 3회 이상 검색 후 helpfulness 평균 < 0.3이면 min_score 요건 상향\\n\\n**변경 파일**:\\n- `src/hooks/stop.ts` — 자동 helpfulness 평가 로직 추가\\n- `src/core/matcher.ts` — helpfulness 가중치 추가\\n- `src/core/graduation.ts` — helpfulness 기반 강등 규칙 추가\\n\\n---\\n\\n### IMP-07: 검색 결과 컨텍스트 포맷 개선\\n\\n**우선순위**: P2\\n\\n**문제**:\\n- 검색된 메모리가 Claude에게 전달될 때의 포맷 불명확\\n- 메모리 출처(세션 날짜, 프로젝트)가 표시되지 않을 수 있음\\n- 100% 선택률 → 품질 필터링 없이 모두 주입\\n\\n**해결 방안**:\\n\\n**A. 메모리 컨텍스트 포맷 구조화**:\\n```\\n<memory source=\\\"2026-02-25\\\" session=\\\"5ef326be\\\" confidence=\\\"0.87\\\">\\n  [쇼핑 어시스턴트 LLM function call 구현 중]\\n  이전에 generalize_with_llm() 함수를 사용해서 상품 추천 로직을 개선했음.\\n  결과: 응답 속도 40% 향상.\\n</memory>\\n```\\n\\n**B. 신뢰도 기준 필터링 강화**:\\n- `high` confidence만 자동 주입 (현재 ≥0.92)\\n- `suggested` confidence는 optional hint로 분리\\n- 100% 선택률 → 선택적 포함으로 변경\\n\\n**변경 파일**:\\n- `src/hooks/user-prompt-submit.ts` — 컨텍스트 포맷 템플릿 개선\\n- `src/core/retriever.ts` — suggested confidence 분리 출력\\n\\n---\\n\\n## 구현 순서\\n\\n```\\nWeek 1 (Critical Path):\\n  IMP-01: Trace 동기 기록 → 즉시 측정 가능\\n  IMP-02: Tool Observation 필터링 → 노이즈 감소\\n\\nWeek 2 (Quality):\\n  IMP-03: 세션 요약 신뢰성\\n  IMP-04: Graduation repair CLI\\n  IMP-05: Embedding 모델 모니터링\\n\\nWeek 3 (Feedback Loop):\\n  IMP-06: Helpfulness 피드백\\n  IMP-07: 컨텍스트 포맷 개선\\n```\\n\\n---\\n\\n## 리스크\\n\\n| 리스크 | 가능성 | 영향 | 대응 |\\n|--------|--------|------|------|\\n| IMP-02로 중요 tool observation 누락 | 중 | 중 | 키워드 필터 화이트리스트 설정 |\\n| IMP-04 graduation repair로 L0 과잉 승격 | 저 | 중 | dry-run 모드 먼저 실행 |\\n| IMP-01 sync 변환으로 hook 응답 지연 | 저 | 저 | trace INSERT는 < 1ms (SQLite sync) |\\n\",\"structuredPatch\":[{\"oldStart\":280,\"oldLines\":20,\"newStart\":280,\"newLines\":73,\"lines\":[\" \",\" ---\",\" \",\"+## 2차 개선 (실측 데이터 기반 후속)\",\"+\",\"+### IMP-01b: 대시보드 API projectId 파라미터 불일치 수정\",\"+\",\"+**문제**: 대시보드가 `?projectId=f4d5c120`으로 쿼리하는데 `getServiceFromQuery()`는 `?project=`만 읽음 → 항상 글로벌 서비스 반환\",\"+\",\"+**수정**: `src/server/api/utils.ts`\",\"+```typescript\",\"+const project = c.req.query('project') || c.req.query('projectId');\",\"+```\",\"+**완료** ✅\",\"+\",\"+---\",\"+\",\"+### IMP-02b: Bash 출력 임계값 800 → 2000\",\"+\",\"+**문제**: 800자 임계값이 너무 낮아 일반적인 bash 출력 대부분이 저장됨\",\"+\",\"+**수정**: `src/hooks/post-tool-use.ts`\",\"+```typescript\",\"+return output.trim().length > 2000;\",\"+```\",\"+**완료** ✅\",\"+\",\"+---\",\"+\",\"+### IMP-06b: Helpfulness 평가 알고리즘 변별력 개선\",\"+\",\"+**문제**: `was_reasked` 로직이 자연스러운 대화 연속을 penalty로 처리 → 모든 항목 0.62~0.68 수렴\",\"+\",\"+**원인**:\",\"+- 쇼핑 어시스턴트처럼 같은 주제를 이어가면 `was_reasked=1` 항상 발생\",\"+- 가중치: `0.20 × (wasReasked ? 0 : 1)` → 모든 항목에 -0.20 penalty\",\"+\",\"+**수정**: `src/core/sqlite-event-store.ts`\",\"+```typescript\",\"+// 기존 (변별력 낮음)\",\"+0.30 × retrievalScore + 0.25 × sessionContinued + 0.25 × toolSuccessRatio + 0.20 × !wasReasked\",\"+\",\"+// 개선 (prompt_count_after 활용)\",\"+0.40 × retrievalScore + 0.30 × promptNorm + 0.20 × toolSuccessRatio + 0.10 × sessionContinued\",\"+// promptNorm = min(promptCountAfter / 2, 1.0)\",\"+// 0턴→0.0, 1턴→0.5, 2턴이상→1.0\",\"+```\",\"+**완료** ✅\",\"+\",\"+---\",\"+\",\" ## 구현 순서\",\" \",\" ```\",\"-Week 1 (Critical Path):\",\"-  IMP-01: Trace 동기 기록 → 즉시 측정 가능\",\"-  IMP-02: Tool Observation 필터링 → 노이즈 감소\",\"+Week 1 (Critical Path) - 완료:\",\"+  IMP-01: Trace projectHash 수정 ✅\",\"+  IMP-02: Tool Observation 필터링 ✅\",\" \",\"-Week 2 (Quality):\",\"-  IMP-03: 세션 요약 신뢰성\",\"+Week 2 (Quality) - 완료:\",\"+  IMP-03: 세션 요약 신뢰성 ✅\",\"   IMP-04: Graduation repair CLI\",\"   IMP-05: Embedding 모델 모니터링\",\" \",\"+Week 2 (후속 실측 기반) - 완료:\",\"+  IMP-01b: 대시보드 projectId 파라미터 수정 ✅\",\"+  IMP-02b: Bash 임계값 800→2000 ✅\",\"+  IMP-06b: Helpfulness 알고리즘 개선 ✅\",\"+\",\" Week 3 (Feedback Loop):\",\"-  IMP-06: Helpfulness 피드백\",\"+  IMP-06: Helpfulness 피드백 루프 완성\",\"   IMP-07: 컨텍스트 포맷 개선\",\" ```\",\" \"]}],\"userModified\":false,\"replaceAll\":false}","durationMs":0,"success":true,"metadata":{"filePath":"/home/walter/workspace/opensource/claude-memory-layer/specs/memory-utilization-improvements/spec.md","fileType":"markdown","turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}
-## 2026-03-04T15:01:18.982Z | c1c31b32-0c38-425e-95a4-1cdbcedbb274
-- type: tool_observation
-- session: 2fe5b2e5-83ea-4778-b09b-5264c550ea86
-{"toolName":"TaskUpdate","toolInput":{"taskId":"3","status":"completed"},"toolOutput":"{\"success\":true,\"taskId\":\"3\",\"updatedFields\":[\"status\"],\"statusChange\":{\"from\":\"in_progress\",\"to\":\"completed\"}}","durationMs":0,"success":true,"metadata":{"turnId":"e42caf45-de79-4d73-a7bc-59954990b764"}}

package/memory/user_prompt/uncategorized/2026-02-26.md

@@ -1,25 +0,0 @@
-
-## 2026-02-26T12:47:51.862Z | 608dc74d-7cda-457c-8455-4cc2b5300fd1
-- type: user_prompt
-- session: 2f324dae-7d8f-4f67-842e-acf507b79c59
-api/stats/retrieval-traces?limit=20
-요청 했을때     "error": "no such table: retrieval_traces"
-## 2026-02-26T12:50:53.044Z | 71b502de-eda8-4d54-a1f4-3ef06fd9edb3
-- type: user_prompt
-- session: 2f324dae-7d8f-4f67-842e-acf507b79c59
-스키마 생성은 언제 하는거야? 
-## 2026-02-26T13:24:02.278Z | dd9328f4-430f-49d7-b8ab-62f39cea6164
-- type: user_prompt
-- session: f8003696-c17e-4e1d-819b-9502d73b9bd4
-suggested_actions 를 생성할때 
-"각 색상·사이즈별 재고 가능 여부 확인해줘"
-같이 없는 기능은 생성하지 않도록 프롬프트 개선해줘 
-
-## 2026-02-26T13:53:54.228Z | efefccef-1a9a-4e89-b1fa-b855fd8fdc85
-- type: user_prompt
-- session: bench-session-2
-최근 변경 사항 때문에 claude code 가 느려진 원인을 분석해줘.
-## 2026-02-26T13:54:45.879Z | f7fca9e1-5169-4732-8845-34fdb5e699c5
-- type: user_prompt
-- session: bench-session-3
-최근 변경 사항 때문에 claude code 가 느려진 원인을 분석해줘.