@onenomad/engram-mcp 1.0.0 → 2.0.0

package/dist/wal.js

@@ -1,296 +1,296 @@
-import { randomUUID } from 'node:crypto';
-import { embed } from './llm.js';
-import { buildContextPrefix } from './utils.js';
-import { chunkContent } from './chunker.js';
-import { extractAndPersistTriples } from './kg-extractor.js';
-import { sourceDedup } from './source-dedup.js';
-// Lightweight poisoning patterns checked at ingest time (no LLM, no search)
-const POISON_PATTERNS = [
-    /\b(ignore previous instructions|ignore all instructions|disregard|forget everything)\b/i,
-    /^(system|SYSTEM)\s*:/m,
-    /\b(act as|you are now|pretend to be|new persona|new identity)\b/i,
-];
-function checkContentPoisoning(content) {
-    for (const pattern of POISON_PATTERNS) {
-        if (pattern.test(content))
-            return 'Suspicious content pattern detected — flagged for review';
-    }
-    return null;
-}
-// ─────────────────────────────────────────────────────────────────────
-// Background side-effect tracking
-// ─────────────────────────────────────────────────────────────────────
-const pendingSideEffects = new Set();
-/**
- * Wait for all in-flight background side-effects (KG extraction +
- * daily-entry append fired with `awaitSideEffects: false`) to
- * complete. No-op when nothing is pending.
- *
- * Tests should call this between ingest and assert; shutdown code
- * should call before process exit to avoid losing KG writes.
- */
-export async function flushPendingSideEffects() {
-    // Snapshot — new promises added during await won't be drained by
-    // this call (they get the next one). Loop until empty in case of
-    // long-running chains.
-    let attempts = 0;
-    while (pendingSideEffects.size > 0 && attempts < 100) {
-        const snapshot = Array.from(pendingSideEffects);
-        await Promise.allSettled(snapshot);
-        attempts++;
-    }
-}
-/** Pending count — for tests + telemetry. */
-export function pendingSideEffectCount() {
-    return pendingSideEffects.size;
-}
-/**
- * Immediately persist one or more memory entries.
- * Designed to be called mid-conversation, before the agent responds.
- */
-export async function ingest(config, storage, entries) {
-    const chunks = [];
-    // Freshly-minted chunks that still need persisting. Cached-source
-    // stubs are added to `chunks` (so callers get them back) but skipped
-    // here, since the underlying rows are already on disk from a prior
-    // ingest. Flushed via storage.saveChunks() in one shot after the
-    // entries loop — replaces N round-trips with 1 against the backend.
-    const newChunks = [];
-    for (const entry of entries) {
-        if (!entry.content || entry.content.trim().length < 5)
-            continue;
-        const trimmedContent = entry.content.trim();
-        // Advisory poisoning check — log warning but never block
-        const poisonFlag = checkContentPoisoning(trimmedContent);
-        if (poisonFlag) {
-            console.error(`Engram governance: ${poisonFlag} in "${trimmedContent.slice(0, 80)}..."`);
-        }
-        // Same-source ingest dedup. When the agent re-reads a stable file
-        // or re-polls an unchanged endpoint within the same Engram process,
-        // we've already chunked + embedded + saved this content. Look up
-        // the (source, content-hash) pair in the in-memory cache and short-
-        // circuit the rest of the pipeline on a hit. Reuses the prior
-        // chunk(s) rather than writing duplicates.
-        //
-        // Bounded session-scoped cache (max 64 sources × 8 hashes); see
-        // source-dedup.ts. Persistence layer doesn't change.
-        const cached = sourceDedup.lookup(entry.source, trimmedContent);
-        if (cached) {
-            // Materialize a chunk reference for the caller from the cached
-            // metadata. We don't re-fetch the actual StoredChunk from disk —
-            // the caller's response only needs id + content + minimal meta,
-            // and the agent's history is keyed off `id`.
-            const stub = {
-                id: cached.chunkId,
-                tier: entry.tier ?? 'short-term',
-                type: entry.type ?? 'context',
-                cognitiveLayer: entry.layer ?? 'episodic',
-                tags: entry.tags ?? [],
-                domain: entry.domain ?? '',
-                topic: entry.topic ?? '',
-                source: entry.source ?? '',
-                importance: entry.importance ?? 0.5,
-                sentiment: entry.sentiment ?? 'neutral',
-                createdAt: new Date().toISOString(),
-                lastRecalledAt: null,
-                recallCount: 0,
-                relatedMemories: [],
-                recallOutcomes: [],
-                origin: entry.origin ?? 'derived',
-                content: trimmedContent,
-            };
-            chunks.push(stub);
-            continue;
-        }
-        const baseType = entry.type ?? inferType(trimmedContent);
-        const baseLayer = entry.layer ?? inferLayer(trimmedContent);
-        // Emotion-weighted importance: high-arousal events get stronger encoding
-        // Matches amygdala research — negative high-arousal memories form faster (0.8 LR)
-        // than positive ones (0.2 LR). Neutral emotions don't modify importance.
-        let effectiveImportance = entry.importance ?? 0.5;
-        if (entry.emotionalArousal !== undefined && entry.emotionalArousal > 0.3) {
-            const valence = entry.emotionalValence ?? 0;
-            // Negative-biased boost: negative emotions boost more than positive
-            const emotionBoost = entry.emotionalArousal * (valence < 0 ? 0.3 : 0.15);
-            effectiveImportance = Math.min(1, effectiveImportance + emotionBoost);
-        }
-        const baseMeta = {
-            tier: entry.tier ?? 'short-term',
-            type: baseType,
-            cognitiveLayer: baseLayer,
-            tags: entry.tags ?? [],
-            domain: entry.domain ?? '',
-            topic: entry.topic ?? '',
-            source: entry.source ?? `wal:${Date.now()}`,
-            importance: effectiveImportance,
-            sentiment: entry.sentiment ?? 'neutral',
-            // Honor caller-provided createdAt (for backfilled memories with
-            // a known original time) — defaults to "now" when omitted.
-            createdAt: entry.createdAt ?? new Date().toISOString(),
-            lastRecalledAt: null,
-            recallCount: 0,
-            relatedMemories: [],
-            recallOutcomes: [],
-            origin: entry.origin ?? 'derived',
-        };
-        // Check if content should be split into sub-chunks
-        const splitResult = config.enableChunking ? chunkContent(trimmedContent) : { chunks: [trimmedContent], needsSplit: false };
-        if (splitResult.needsSplit) {
-            // Save parent chunk (no embedding, used for keyword search)
-            const parentChunk = {
-                id: randomUUID(),
-                ...baseMeta,
-                content: trimmedContent,
-                consolidationLevel: -1, // Sentinel: parent container
-            };
-            newChunks.push(parentChunk);
-            chunks.push(parentChunk);
-            // Remember the parent chunk id keyed by source so a re-ingest
-            // of the identical content within the same process returns this
-            // same id and skips chunk+embed+save entirely.
-            sourceDedup.remember(entry.source, trimmedContent, parentChunk.id);
-            // Save sub-chunks with embeddings
-            for (const subContent of splitResult.chunks) {
-                const subChunk = {
-                    id: randomUUID(),
-                    ...baseMeta,
-                    content: subContent,
-                    parentChunkId: parentChunk.id,
-                };
-                // Detect temporal anchor
-                const dateMatch = subContent.match(/\b(\d{4})-(\d{2})-(\d{2})\b/) ??
-                    subContent.match(/\b(january|february|march|april|may|june|july|august|september|october|november|december)\s+(\d{1,2})[,.]?\s+(\d{4})\b/i);
-                if (dateMatch) {
-                    try {
-                        const parsed = new Date(dateMatch[0]);
-                        if (!isNaN(parsed.getTime()))
-                            subChunk.temporalAnchor = parsed.getTime();
-                    }
-                    catch { /* skip */ }
-                }
-                try {
-                    const prefix = buildContextPrefix(subChunk);
-                    subChunk.embedding = await embed(config, subContent, prefix);
-                    subChunk.embeddingVersion = 1;
-                }
-                catch { /* skip */ }
-                newChunks.push(subChunk);
-                chunks.push(subChunk);
-            }
-        }
-        else {
-            // Single chunk path (original behavior)
-            const chunk = {
-                id: randomUUID(),
-                ...baseMeta,
-                content: trimmedContent,
-            };
-            const dateMatch = chunk.content.match(/\b(\d{4})-(\d{2})-(\d{2})\b/) ??
-                chunk.content.match(/\b(january|february|march|april|may|june|july|august|september|october|november|december)\s+(\d{1,2})[,.]?\s+(\d{4})\b/i);
-            if (dateMatch) {
-                try {
-                    const parsed = new Date(dateMatch[0]);
-                    if (!isNaN(parsed.getTime()))
-                        chunk.temporalAnchor = parsed.getTime();
-                }
-                catch { /* skip */ }
-            }
-            try {
-                const prefix = buildContextPrefix(chunk);
-                chunk.embedding = await embed(config, chunk.content, prefix);
-                chunk.embeddingVersion = 1;
-            }
-            catch { /* skip */ }
-            newChunks.push(chunk);
-            chunks.push(chunk);
-            // Single-chunk path: cache the chunk id keyed by source.
-            sourceDedup.remember(entry.source, trimmedContent, chunk.id);
-        }
-    }
-    // One batched write for every new chunk in the call.
-    if (newChunks.length > 0) {
-        await storage.saveChunks(newChunks);
-    }
-    // Per-batch side effects. Both opt-out via flags on any entry in
-    // the batch (typical: memory_ingest calls ingest() with one entry,
-    // so a single flag controls the path). Benchmark harnesses set
-    // these to match what engram/benchmarks/locomo.ts does — its
-    // direct-saveChunk path skips both, which is the source of the
-    // ~50× wall-clock gap between standalone and MCP-boundary benches.
-    const skipDaily = entries.some(e => e.skipDailyEntry);
-    const skipKg = entries.some(e => e.skipKgExtraction);
-    // awaitSideEffects defaults TRUE — only flip to async when EVERY
-    // entry in the batch opts out, to avoid surprising a sync caller
-    // batched with an async one.
-    const runAsync = entries.length > 0 && entries.every(e => e.awaitSideEffects === false);
-    if (chunks.length > 0) {
-        const sideEffectsTask = async () => {
-            if (!skipDaily) {
-                const date = new Date().toISOString().split('T')[0];
-                try {
-                    await storage.appendDailyEntry(date, {
-                        timestamp: new Date().toISOString(),
-                        conversationId: chunks[0].source,
-                        summary: `WAL ingest: ${chunks.length} entries`,
-                        extractedFacts: chunks.map(c => c.content),
-                    });
-                }
-                catch {
-                    // best-effort: a daily-entry append failure must not break
-                    // the rest of the side-effects task
-                }
-            }
-            if (!skipKg) {
-                // Auto-populate knowledge graph from ingested content
-                for (const chunk of chunks) {
-                    if (chunk.consolidationLevel === -1)
-                        continue; // skip parent containers
-                    try {
-                        await extractAndPersistTriples(storage, chunk.content, {
-                            domain: chunk.domain,
-                            topic: chunk.topic,
-                            source: chunk.source,
-                        });
-                    }
-                    catch {
-                        // KG extraction is best-effort — never block ingestion
-                    }
-                }
-            }
-        };
-        if (runAsync) {
-            // Fire and forget — track in pendingSideEffects so tests or
-            // shutdown code can drain via flushPendingSideEffects().
-            const p = sideEffectsTask()
-                .catch(() => { })
-                .finally(() => { pendingSideEffects.delete(p); });
-            pendingSideEffects.add(p);
-        }
-        else {
-            await sideEffectsTask();
-        }
-    }
-    return chunks;
-}
-// ── Type/Layer inference heuristics ──────────────────────────────────
-function inferType(content) {
-    const lower = content.toLowerCase();
-    if (lower.includes('prefer') || lower.includes('like') || lower.includes('want'))
-        return 'preference';
-    if (lower.includes('decided') || lower.includes('going with') || lower.includes('chose') || lower.includes('use '))
-        return 'decision';
-    if (lower.includes('not ') || lower.includes('wrong') || lower.includes('correct') || lower.includes('instead'))
-        return 'correction';
-    if (lower.includes('working on') || lower.includes('currently') || lower.includes('right now'))
-        return 'context';
-    return 'fact';
-}
-function inferLayer(content) {
-    const lower = content.toLowerCase();
-    if (lower.includes('always') || lower.includes('never') || lower.includes('rule') || lower.includes('should'))
-        return 'procedural';
-    if (lower.includes('today') || lower.includes('yesterday') || lower.includes('just ') || lower.includes('session'))
-        return 'episodic';
-    return 'semantic';
-}
+import { randomUUID } from 'node:crypto';
+import { embed } from './llm.js';
+import { buildContextPrefix } from './utils.js';
+import { chunkContent } from './chunker.js';
+import { extractAndPersistTriples } from './kg-extractor.js';
+import { sourceDedup } from './source-dedup.js';
+// Lightweight poisoning patterns checked at ingest time (no LLM, no search)
+const POISON_PATTERNS = [
+    /\b(ignore previous instructions|ignore all instructions|disregard|forget everything)\b/i,
+    /^(system|SYSTEM)\s*:/m,
+    /\b(act as|you are now|pretend to be|new persona|new identity)\b/i,
+];
+function checkContentPoisoning(content) {
+    for (const pattern of POISON_PATTERNS) {
+        if (pattern.test(content))
+            return 'Suspicious content pattern detected — flagged for review';
+    }
+    return null;
+}
+// ─────────────────────────────────────────────────────────────────────
+// Background side-effect tracking
+// ─────────────────────────────────────────────────────────────────────
+const pendingSideEffects = new Set();
+/**
+ * Wait for all in-flight background side-effects (KG extraction +
+ * daily-entry append fired with `awaitSideEffects: false`) to
+ * complete. No-op when nothing is pending.
+ *
+ * Tests should call this between ingest and assert; shutdown code
+ * should call before process exit to avoid losing KG writes.
+ */
+export async function flushPendingSideEffects() {
+    // Snapshot — new promises added during await won't be drained by
+    // this call (they get the next one). Loop until empty in case of
+    // long-running chains.
+    let attempts = 0;
+    while (pendingSideEffects.size > 0 && attempts < 100) {
+        const snapshot = Array.from(pendingSideEffects);
+        await Promise.allSettled(snapshot);
+        attempts++;
+    }
+}
+/** Pending count — for tests + telemetry. */
+export function pendingSideEffectCount() {
+    return pendingSideEffects.size;
+}
+/**
+ * Immediately persist one or more memory entries.
+ * Designed to be called mid-conversation, before the agent responds.
+ */
+export async function ingest(config, storage, entries) {
+    const chunks = [];
+    // Freshly-minted chunks that still need persisting. Cached-source
+    // stubs are added to `chunks` (so callers get them back) but skipped
+    // here, since the underlying rows are already on disk from a prior
+    // ingest. Flushed via storage.saveChunks() in one shot after the
+    // entries loop — replaces N round-trips with 1 against the backend.
+    const newChunks = [];
+    for (const entry of entries) {
+        if (!entry.content || entry.content.trim().length < 5)
+            continue;
+        const trimmedContent = entry.content.trim();
+        // Advisory poisoning check — log warning but never block
+        const poisonFlag = checkContentPoisoning(trimmedContent);
+        if (poisonFlag) {
+            console.error(`Engram governance: ${poisonFlag} in "${trimmedContent.slice(0, 80)}..."`);
+        }
+        // Same-source ingest dedup. When the agent re-reads a stable file
+        // or re-polls an unchanged endpoint within the same Engram process,
+        // we've already chunked + embedded + saved this content. Look up
+        // the (source, content-hash) pair in the in-memory cache and short-
+        // circuit the rest of the pipeline on a hit. Reuses the prior
+        // chunk(s) rather than writing duplicates.
+        //
+        // Bounded session-scoped cache (max 64 sources × 8 hashes); see
+        // source-dedup.ts. Persistence layer doesn't change.
+        const cached = sourceDedup.lookup(entry.source, trimmedContent);
+        if (cached) {
+            // Materialize a chunk reference for the caller from the cached
+            // metadata. We don't re-fetch the actual StoredChunk from disk —
+            // the caller's response only needs id + content + minimal meta,
+            // and the agent's history is keyed off `id`.
+            const stub = {
+                id: cached.chunkId,
+                tier: entry.tier ?? 'short-term',
+                type: entry.type ?? 'context',
+                cognitiveLayer: entry.layer ?? 'episodic',
+                tags: entry.tags ?? [],
+                domain: entry.domain ?? '',
+                topic: entry.topic ?? '',
+                source: entry.source ?? '',
+                importance: entry.importance ?? 0.5,
+                sentiment: entry.sentiment ?? 'neutral',
+                createdAt: new Date().toISOString(),
+                lastRecalledAt: null,
+                recallCount: 0,
+                relatedMemories: [],
+                recallOutcomes: [],
+                origin: entry.origin ?? 'derived',
+                content: trimmedContent,
+            };
+            chunks.push(stub);
+            continue;
+        }
+        const baseType = entry.type ?? inferType(trimmedContent);
+        const baseLayer = entry.layer ?? inferLayer(trimmedContent);
+        // Emotion-weighted importance: high-arousal events get stronger encoding
+        // Matches amygdala research — negative high-arousal memories form faster (0.8 LR)
+        // than positive ones (0.2 LR). Neutral emotions don't modify importance.
+        let effectiveImportance = entry.importance ?? 0.5;
+        if (entry.emotionalArousal !== undefined && entry.emotionalArousal > 0.3) {
+            const valence = entry.emotionalValence ?? 0;
+            // Negative-biased boost: negative emotions boost more than positive
+            const emotionBoost = entry.emotionalArousal * (valence < 0 ? 0.3 : 0.15);
+            effectiveImportance = Math.min(1, effectiveImportance + emotionBoost);
+        }
+        const baseMeta = {
+            tier: entry.tier ?? 'short-term',
+            type: baseType,
+            cognitiveLayer: baseLayer,
+            tags: entry.tags ?? [],
+            domain: entry.domain ?? '',
+            topic: entry.topic ?? '',
+            source: entry.source ?? `wal:${Date.now()}`,
+            importance: effectiveImportance,
+            sentiment: entry.sentiment ?? 'neutral',
+            // Honor caller-provided createdAt (for backfilled memories with
+            // a known original time) — defaults to "now" when omitted.
+            createdAt: entry.createdAt ?? new Date().toISOString(),
+            lastRecalledAt: null,
+            recallCount: 0,
+            relatedMemories: [],
+            recallOutcomes: [],
+            origin: entry.origin ?? 'derived',
+        };
+        // Check if content should be split into sub-chunks
+        const splitResult = config.enableChunking ? chunkContent(trimmedContent) : { chunks: [trimmedContent], needsSplit: false };
+        if (splitResult.needsSplit) {
+            // Save parent chunk (no embedding, used for keyword search)
+            const parentChunk = {
+                id: randomUUID(),
+                ...baseMeta,
+                content: trimmedContent,
+                consolidationLevel: -1, // Sentinel: parent container
+            };
+            newChunks.push(parentChunk);
+            chunks.push(parentChunk);
+            // Remember the parent chunk id keyed by source so a re-ingest
+            // of the identical content within the same process returns this
+            // same id and skips chunk+embed+save entirely.
+            sourceDedup.remember(entry.source, trimmedContent, parentChunk.id);
+            // Save sub-chunks with embeddings
+            for (const subContent of splitResult.chunks) {
+                const subChunk = {
+                    id: randomUUID(),
+                    ...baseMeta,
+                    content: subContent,
+                    parentChunkId: parentChunk.id,
+                };
+                // Detect temporal anchor
+                const dateMatch = subContent.match(/\b(\d{4})-(\d{2})-(\d{2})\b/) ??
+                    subContent.match(/\b(january|february|march|april|may|june|july|august|september|october|november|december)\s+(\d{1,2})[,.]?\s+(\d{4})\b/i);
+                if (dateMatch) {
+                    try {
+                        const parsed = new Date(dateMatch[0]);
+                        if (!isNaN(parsed.getTime()))
+                            subChunk.temporalAnchor = parsed.getTime();
+                    }
+                    catch { /* skip */ }
+                }
+                try {
+                    const prefix = buildContextPrefix(subChunk);
+                    subChunk.embedding = await embed(config, subContent, prefix);
+                    subChunk.embeddingVersion = 1;
+                }
+                catch { /* skip */ }
+                newChunks.push(subChunk);
+                chunks.push(subChunk);
+            }
+        }
+        else {
+            // Single chunk path (original behavior)
+            const chunk = {
+                id: randomUUID(),
+                ...baseMeta,
+                content: trimmedContent,
+            };
+            const dateMatch = chunk.content.match(/\b(\d{4})-(\d{2})-(\d{2})\b/) ??
+                chunk.content.match(/\b(january|february|march|april|may|june|july|august|september|october|november|december)\s+(\d{1,2})[,.]?\s+(\d{4})\b/i);
+            if (dateMatch) {
+                try {
+                    const parsed = new Date(dateMatch[0]);
+                    if (!isNaN(parsed.getTime()))
+                        chunk.temporalAnchor = parsed.getTime();
+                }
+                catch { /* skip */ }
+            }
+            try {
+                const prefix = buildContextPrefix(chunk);
+                chunk.embedding = await embed(config, chunk.content, prefix);
+                chunk.embeddingVersion = 1;
+            }
+            catch { /* skip */ }
+            newChunks.push(chunk);
+            chunks.push(chunk);
+            // Single-chunk path: cache the chunk id keyed by source.
+            sourceDedup.remember(entry.source, trimmedContent, chunk.id);
+        }
+    }
+    // One batched write for every new chunk in the call.
+    if (newChunks.length > 0) {
+        await storage.saveChunks(newChunks);
+    }
+    // Per-batch side effects. Both opt-out via flags on any entry in
+    // the batch (typical: engram-ingest calls ingest() with one entry,
+    // so a single flag controls the path). Benchmark harnesses set
+    // these to match what engram/benchmarks/locomo.ts does — its
+    // direct-saveChunk path skips both, which is the source of the
+    // ~50× wall-clock gap between standalone and MCP-boundary benches.
+    const skipDaily = entries.some(e => e.skipDailyEntry);
+    const skipKg = entries.some(e => e.skipKgExtraction);
+    // awaitSideEffects defaults TRUE — only flip to async when EVERY
+    // entry in the batch opts out, to avoid surprising a sync caller
+    // batched with an async one.
+    const runAsync = entries.length > 0 && entries.every(e => e.awaitSideEffects === false);
+    if (chunks.length > 0) {
+        const sideEffectsTask = async () => {
+            if (!skipDaily) {
+                const date = new Date().toISOString().split('T')[0];
+                try {
+                    await storage.appendDailyEntry(date, {
+                        timestamp: new Date().toISOString(),
+                        conversationId: chunks[0].source,
+                        summary: `WAL ingest: ${chunks.length} entries`,
+                        extractedFacts: chunks.map(c => c.content),
+                    });
+                }
+                catch {
+                    // best-effort: a daily-entry append failure must not break
+                    // the rest of the side-effects task
+                }
+            }
+            if (!skipKg) {
+                // Auto-populate knowledge graph from ingested content
+                for (const chunk of chunks) {
+                    if (chunk.consolidationLevel === -1)
+                        continue; // skip parent containers
+                    try {
+                        await extractAndPersistTriples(storage, chunk.content, {
+                            domain: chunk.domain,
+                            topic: chunk.topic,
+                            source: chunk.source,
+                        });
+                    }
+                    catch {
+                        // KG extraction is best-effort — never block ingestion
+                    }
+                }
+            }
+        };
+        if (runAsync) {
+            // Fire and forget — track in pendingSideEffects so tests or
+            // shutdown code can drain via flushPendingSideEffects().
+            const p = sideEffectsTask()
+                .catch(() => { })
+                .finally(() => { pendingSideEffects.delete(p); });
+            pendingSideEffects.add(p);
+        }
+        else {
+            await sideEffectsTask();
+        }
+    }
+    return chunks;
+}
+// ── Type/Layer inference heuristics ──────────────────────────────────
+function inferType(content) {
+    const lower = content.toLowerCase();
+    if (lower.includes('prefer') || lower.includes('like') || lower.includes('want'))
+        return 'preference';
+    if (lower.includes('decided') || lower.includes('going with') || lower.includes('chose') || lower.includes('use '))
+        return 'decision';
+    if (lower.includes('not ') || lower.includes('wrong') || lower.includes('correct') || lower.includes('instead'))
+        return 'correction';
+    if (lower.includes('working on') || lower.includes('currently') || lower.includes('right now'))
+        return 'context';
+    return 'fact';
+}
+function inferLayer(content) {
+    const lower = content.toLowerCase();
+    if (lower.includes('always') || lower.includes('never') || lower.includes('rule') || lower.includes('should'))
+        return 'procedural';
+    if (lower.includes('today') || lower.includes('yesterday') || lower.includes('just ') || lower.includes('session'))
+        return 'episodic';
+    return 'semantic';
+}
 //# sourceMappingURL=wal.js.map