@oh-my-pi/pi-mnemopi 18.18.0 → 18.18.2

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [18.18.2] - 2026-11-08
+
+### Fixed
+
+- Fixed recall returning episodic summaries for working memories that had been invalidated.
+
+## [18.18.1] - 2026-11-08
+
+### Fixed
+
+- Fixed recall scoring memories against vectors from a previous embedding model while a model-change rebuild was still running or had failed: dense search now only uses vectors stamped with the active model, falling back to lexical matching for memories awaiting re-embedding ([#18442](https://github.com/can1357/oh-my-pi/pull/18442) by [@eggpeat](https://github.com/eggpeat))
+- Fixed model-change embedding rebuilds overwriting newer vectors for updated memories or recreating vectors for deleted memories ([#18441](https://github.com/can1357/oh-my-pi/pull/18441) by [@eggpeat](https://github.com/eggpeat))
+
 ## [18.18.0] - 2026-11-07
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-mnemopi",
-	"version": "18.18.0",
+	"version": "18.18.2",
 	"description": "Local SQLite memory engine for omp agents",
 	"homepage": "https://omp.sh",
 	"author": {
@@ -42,10 +42,10 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "18.18.0",
-		"@oh-my-pi/pi-catalog": "18.18.0",
-		"@oh-my-pi/pi-natives": "18.18.0",
-		"@oh-my-pi/pi-utils": "18.18.0"
+		"@oh-my-pi/pi-ai": "18.18.2",
+		"@oh-my-pi/pi-catalog": "18.18.2",
+		"@oh-my-pi/pi-natives": "18.18.2",
+		"@oh-my-pi/pi-utils": "18.18.2"
 	},
 	"peerDependencies": {
 		"fastembed": "3.0.0",

package/src/core/beam/recall.ts

@@ -13,7 +13,7 @@ import {
 	unknownWeight,
 } from "../../config";
 import { hasCjk, matchesWordForm } from "../../util/regex";
-import { embedQuery } from "../embeddings";
+import { currentEmbeddingModel, embedQuery } from "../embeddings";
 import { mmrRerank } from "../mmr";
 import { type OrchestratedRecallResult, orchestrateRecall } from "../orchestrator";
 import { POLYPHONIC_MAX_COMBINED_SCORE } from "../polyphonic-recall";
@@ -604,12 +604,15 @@ function vectorSimilarities(
 	) {
 		return out;
 	}
+	// Only vectors stamped with the active embedding model are comparable to the
+	// query embedding; stale-model rows stay on disk for the rebuild but must not score.
+	const model = currentEmbeddingModel();
 	for (let offset = 0; offset < memoryIds.length; offset += 500) {
 		const chunk = memoryIds.slice(offset, offset + 500);
 		const rows = queryAll(
 			beam,
-			`SELECT memory_id, embedding_json FROM memory_embeddings WHERE memory_id IN (${placeholders(chunk.length)})`,
-			chunk,
+			`SELECT memory_id, embedding_json FROM memory_embeddings WHERE model IS ? AND memory_id IN (${placeholders(chunk.length)})`,
+			[model, ...chunk],
 		);
 		for (const row of rows) {
 			const vector = parseEmbedding(row.embedding_json);

package/src/core/beam/store.ts

@@ -407,7 +407,7 @@ export function reconcileEmbeddingModel(beam: BeamMemoryState): void {
 				// Embed the whole corpus before touching anything stored. embed() handles the provider's
 				// per-request cap; a rejected chunk abandons the rebuild and leaves the stored vectors alone.
 				const toEmbed = live.filter(item => item.content.trim() !== "");
-				const staged: { memoryId: string; json: string }[] = [];
+				const staged: { memoryId: string; content: string; json: string }[] = [];
 				for (let offset = 0; offset < toEmbed.length; offset += EMBED_REBUILD_BATCH) {
 					const chunk = toEmbed.slice(offset, offset + EMBED_REBUILD_BATCH);
 					const vectors = await embed(chunk.map(item => item.content));
@@ -422,14 +422,18 @@ export function reconcileEmbeddingModel(beam: BeamMemoryState): void {
 					chunk.forEach((item, index) => {
 						const vector = vectors[index];
 						if (vector !== undefined) {
-							staged.push({ memoryId: item.memoryId, json: JSON.stringify(Array.from(vector)) });
+							staged.push({
+								memoryId: item.memoryId,
+								content: item.content,
+								json: JSON.stringify(Array.from(vector)),
+							});
 						}
 					});
 				}
 
 				transaction(beam.db, () => {
-					// Only stale-model rows go. A vector that a concurrent remember() stamped under the active
-					// model survives, and the INSERT below overwrites any row this rebuild covers.
+					// Only stale-model rows go. A vector that a concurrent remember() or update() stamped
+					// under the active model survives. Each staged insert also revalidates its source text.
 					beam.db.query("DELETE FROM memory_embeddings WHERE model IS NOT ?").run(active);
 					beam.db.run("UPDATE episodic_memory SET binary_vector = NULL");
 					if (vecAvailable(beam.db)) {
@@ -439,10 +443,28 @@ export function reconcileEmbeddingModel(beam: BeamMemoryState): void {
 							// sqlite-vec cleanup is best-effort; rebuild correctness takes precedence.
 						}
 					}
-					using insertEmbedding = beam.db.prepare(
-						"INSERT OR REPLACE INTO memory_embeddings(memory_id, embedding_json, model) VALUES (?, ?, ?)",
-					);
-					for (const row of staged) insertEmbedding.run(row.memoryId, row.json, active);
+					using insertEmbedding = beam.db.prepare(`
+						INSERT OR REPLACE INTO memory_embeddings(memory_id, embedding_json, model)
+						SELECT ?, ?, ?
+						WHERE EXISTS (
+							SELECT 1 FROM working_memory
+							WHERE id = ? AND superseded_by IS NULL AND COALESCE(embed_text, content) = ?
+						) OR EXISTS (
+							SELECT 1 FROM episodic_memory
+							WHERE id = ? AND superseded_by IS NULL AND content = ?
+						)
+					`);
+					for (const row of staged) {
+						insertEmbedding.run(
+							row.memoryId,
+							row.json,
+							active,
+							row.memoryId,
+							row.content,
+							row.memoryId,
+							row.content,
+						);
+					}
 				});
 				beam.caches.queryCache?.invalidate();
 
@@ -693,37 +715,39 @@ export function getContext(beam: BeamMemoryState, limit = 10): Row[] {
 
 export function invalidate(beam: BeamMemoryState, memoryId: string, replacementId: string | null = null): boolean {
 	const now = toUtcIso();
-	const working = beam.db.run(
-		`
-			UPDATE working_memory
-			SET valid_until = ?, superseded_by = ?
-			WHERE id = ? AND (session_id = ? OR scope = 'global')
-		`,
-		[now, replacementId, memoryId, beam.sessionId],
-	);
-	if (working.changes > 0) {
-		// Recall filters `valid_until`/`superseded_by` in SQL, but the enhanced path consults the
-		// query cache BEFORE it reaches SQL. Without this the row a caller just retired keeps being
-		// served to an identical query for the rest of the cache TTL -- the one thing an explicit
-		// invalidation is supposed to guarantee against. Every other mutator here already does this;
-		// this one was the omission. Gated on an actual row change, like `forgetWorking`, so a
-		// no-op invalidation never discards a valid cache.
-		invalidateCaches(beam);
-		return true;
-	}
-	const episodic = beam.db.run(
-		`
-			UPDATE episodic_memory
-			SET valid_until = ?, superseded_by = ?
-			WHERE id = ? AND (session_id = ? OR scope = 'global')
-		`,
-		[now, replacementId, memoryId, beam.sessionId],
-	);
-	if (episodic.changes > 0) {
-		invalidateCaches(beam);
-		return true;
-	}
-	return false;
+	const changed = transaction(beam.db, () => {
+		const working = beam.db.run(
+			`
+				UPDATE working_memory
+				SET valid_until = ?, superseded_by = ?
+				WHERE id = ? AND (session_id = ? OR scope = 'global')
+			`,
+			[now, replacementId, memoryId, beam.sessionId],
+		);
+		if (working.changes > 0) {
+			// Retire episodic summaries built from this source: they still carry its text. Siblings
+			// covered by the same summary stay recallable via their (consolidated, never TTL-trimmed)
+			// working rows.
+			beam.db.run(
+				`UPDATE episodic_memory SET valid_until = ?, superseded_by = ?
+				 WHERE (session_id = ? OR scope = 'global')
+				   AND instr(',' || summary_of || ',', ',' || ? || ',') > 0`,
+				[now, replacementId, beam.sessionId, memoryId],
+			);
+			return true;
+		}
+		const episodic = beam.db.run(
+			`
+				UPDATE episodic_memory
+				SET valid_until = ?, superseded_by = ?
+				WHERE id = ? AND (session_id = ? OR scope = 'global')
+			`,
+			[now, replacementId, memoryId, beam.sessionId],
+		);
+		return episodic.changes > 0;
+	});
+	if (changed) invalidateCaches(beam);
+	return changed;
 }
 
 export function getWorkingStats(

package/src/core/polyphonic-recall.ts

@@ -5,6 +5,7 @@ import { closeQuietly, type DatabasePath, openDatabase } from "../db";
 import { backfillConsolidatedFacts, ensureVeracityConsolidator } from "./beam/consolidate";
 import type { BeamMemoryState, JsonValue, Metadata, RecallResult, RecallTierLabel } from "./beam/types";
 import { EpisodicGraph, extractProperNouns } from "./episodic-graph";
+import { currentEmbeddingModel } from "./embeddings";
 import { STOP_WORDS } from "./synonyms";
 import { VeracityConsolidator } from "./veracity-consolidation";
 
@@ -342,6 +343,7 @@ export class PolyphonicRecallEngine {
 		const queryUnit = normalizeVector(queryEmbedding);
 		if (queryUnit === null) return [];
 		const now = new Date().toISOString();
+		const model = currentEmbeddingModel();
 		const working = this.#visibility("wm", options.channelId);
 		const episodic = this.#visibility("em", options.channelId);
 		let rows: EmbeddingRow[] = [];
@@ -351,19 +353,21 @@ export class PolyphonicRecallEngine {
 					SELECT me.memory_id, me.embedding_json, 'working' AS embedding_tier
 					FROM memory_embeddings me
 					JOIN working_memory wm ON wm.id = me.memory_id
-					WHERE wm.superseded_by IS NULL
+					WHERE me.model IS ?
+						AND wm.superseded_by IS NULL
 						AND (wm.valid_until IS NULL OR wm.valid_until > ?)
 						AND ${working.clause}
 					UNION ALL
 					SELECT me.memory_id, me.embedding_json, 'episodic' AS embedding_tier
 					FROM memory_embeddings me
 					JOIN episodic_memory em ON em.id = me.memory_id
-					WHERE em.superseded_by IS NULL
+					WHERE me.model IS ?
+						AND em.superseded_by IS NULL
 						AND (em.valid_until IS NULL OR em.valid_until > ?)
 						AND ${episodic.clause}
 					LIMIT 50000
 				`)
-				.all(now, ...working.params, now, ...episodic.params) as EmbeddingRow[];
+				.all(model, now, ...working.params, model, now, ...episodic.params) as EmbeddingRow[];
 		} catch {
 			return [];
 		}