@prometheus-ai/memory 0.5.3 → 0.5.8

package/src/core/beam/recall.ts

@@ -2,7 +2,7 @@ import { normalizedRecallWeights, temporalHalflifeHours } from "../../config";
 import { embedQuery } from "../embeddings";
 import { mmrRerank } from "../mmr";
 import { adjustWeights, classifyIntent } from "../query-intent";
-import { getSynonyms, normalizeQuery } from "../synonyms";
+import { getSynonyms, normalizeQuery, STOP_WORDS as QUERY_STOP_WORDS } from "../synonyms";
 import { extractTemporal } from "../temporal-parser";
 import { cosineSimilarity } from "../vector-math";
 import type { BeamMemoryState, RecallEnhancedOptions, RecallOptions, RecallResult } from "./types";
@@ -101,6 +101,31 @@ const STOP_WORDS = new Set([
 	"with",
 ]);
 
+const FACT_QUERY_FILLER_WORDS = new Set([
+	...QUERY_STOP_WORDS,
+	"active",
+	"current",
+	"currently",
+	"d",
+	"know",
+	"latest",
+	"ll",
+	"m",
+	"please",
+	"present",
+	"re",
+	"recent",
+	"remind",
+	"remember",
+	"s",
+	"t",
+	"tell",
+	"today",
+	"ve",
+]);
+
+const FACT_CLITIC_FRAGMENTS = new Set(["d", "ll", "m", "re", "s", "t", "ve"]);
+
 function nowIso(): string {
 	return new Date().toISOString();
 }
@@ -176,6 +201,35 @@ function expandedTokenGroups(query: string, useSynonyms = true): string[][] {
 	return groups;
 }
 
+function factExpandedTokenGroups(query: string, content: string): string[][] {
+	const contentLower = content.toLowerCase();
+	const contentTokens = new Set(tokenize(contentLower));
+	const groups: string[][] = [];
+	for (const token of tokenize(query)) {
+		if (FACT_QUERY_FILLER_WORDS.has(token) && (FACT_CLITIC_FRAGMENTS.has(token) || !contentTokens.has(token))) {
+			continue;
+		}
+		const seen = new Set<string>();
+		for (const variant of recallSynonyms(token, true)) {
+			for (const part of tokenize(variant)) {
+				if (!FACT_QUERY_FILLER_WORDS.has(part) || (!FACT_CLITIC_FRAGMENTS.has(part) && contentTokens.has(part))) {
+					seen.add(part);
+				}
+			}
+		}
+		if (seen.size > 0) groups.push([...seen]);
+	}
+	return groups;
+}
+
+function tokensFromGroups(groups: readonly (readonly string[])[]): string[] {
+	const seen = new Set<string>();
+	for (const group of groups) {
+		for (const token of group) seen.add(token);
+	}
+	return [...seen];
+}
+
 function contentMatchesToken(contentLower: string, contentTokens: ReadonlySet<string>, token: string): boolean {
 	if (contentTokens.has(token) || contentLower.includes(token)) return true;
 	for (const contentToken of contentTokens) {
@@ -1062,13 +1116,11 @@ export function factRecall(beam: BeamMemoryState, query: string, topK = 30): Fac
 		}
 	}
 	if (matched.length === 0) return [];
-	const rowids = matched
-		.slice(0, topK)
-		.map(row => asNumber(row.rowid))
-		.filter(rowid => rowid > 0);
+	const rowids = matched.map(row => asNumber(row.rowid)).filter(rowid => rowid > 0);
 	if (rowids.length === 0) return [];
 	const visibility = factVisibilityWhere(beam, "");
 	const ranks = normalizeRanks(matched, "rowid");
+	const normalized = normalizeQuery(query).toLowerCase();
 	const rows = queryAll(
 		beam,
 		`SELECT rowid, fact_id, subject, predicate, object, timestamp, confidence
@@ -1076,25 +1128,46 @@ export function factRecall(beam: BeamMemoryState, query: string, topK = 30): Fac
 		 WHERE rowid IN (${placeholders(rowids.length)}) AND ${visibility.where}
 		 ORDER BY confidence DESC
 		 LIMIT ?`,
-		[...rowids, ...visibility.params, topK],
+		[...rowids, ...visibility.params, rowids.length],
 	);
-	return rows.map(row => {
-		const subject = asString(row.subject);
-		const predicate = asString(row.predicate);
-		const object = asString(row.object);
-		const confidence = asNumber(row.confidence, 0.5);
-		const result: FactRecallResult = {
-			id: asString(row.fact_id),
-			content: object.length > 0 ? object : `${subject} ${predicate}`.trim(),
-			score: round4(confidence * 0.8 + (ranks.get(asNumber(row.rowid)) ?? 0) * 0.2),
-			fact_id: asString(row.fact_id),
-			subject,
-			predicate,
-			timestamp: asNullableString(row.timestamp),
-			tier_label: "fact",
-			tier: "fact",
-			source: "facts",
-		};
-		return result;
-	});
+	return rows
+		.map(row => {
+			const subject = asString(row.subject);
+			const predicate = asString(row.predicate);
+			const object = asString(row.object);
+			const confidence = asNumber(row.confidence, 0.5);
+			const content = object.length > 0 ? object : `${subject} ${predicate}`.trim();
+			const searchable = `${subject} ${predicate} ${object}`.trim();
+			const queryGroups = factExpandedTokenGroups(query, searchable);
+			const queryTokens = tokensFromGroups(queryGroups);
+			const lexical =
+				queryGroups.length > 0
+					? lexicalGroupRelevance(queryGroups, searchable, normalized)
+					: lexicalRelevance(queryTokens, searchable, normalized);
+			const rank = ranks.get(asNumber(row.rowid)) ?? 0;
+			const result: FactRecallResult = {
+				id: asString(row.fact_id),
+				content,
+				score: round4(lexical * (0.7 + confidence * 0.2 + rank * 0.1)),
+				fact_id: asString(row.fact_id),
+				subject,
+				predicate,
+				timestamp: asNullableString(row.timestamp),
+				tier_label: "fact",
+				tier: "fact",
+				source: "facts",
+				keyword_score: round4(lexical),
+				fts_score: round4(rank),
+				importance_score: round4(confidence),
+				explanation: `fact keyword=${round4(lexical)}`,
+				voice_scores: {
+					keyword: round4(lexical),
+					fts: round4(rank),
+					importance: round4(confidence),
+				},
+			};
+			return result;
+		})
+		.sort((left, right) => (right.score ?? 0) - (left.score ?? 0))
+		.slice(0, topK);
 }

package/src/core/beam/store.ts

@@ -1,12 +1,14 @@
 import type { Database, SQLQueryBindings } from "bun:sqlite";
+import { logger } from "@prometheus-ai/utils";
 import { transaction } from "../../db";
 import { toUtcIso } from "../../util/datetime";
 import { generateId } from "../../util/ids";
+import { currentEmbeddingModel, embeddingsDisabled } from "../embeddings";
 import { EpisodicGraph } from "../episodic-graph";
 import { extractFactsSafe } from "../extraction";
 import { getMnemopiRuntimeOptions, withMnemopiRuntimeOptions } from "../runtime-options";
 import { storeFactStrings } from "./consolidate";
-import { scheduleEmbedding, vecAvailable, vecInsert } from "./helpers";
+import { type EmbedItem, scheduleEmbedding, vecAvailable, vecInsert } from "./helpers";
 import type {
 	BeamEvent,
 	BeamMemoryState,
@@ -58,7 +60,7 @@ const TRUST_TIERS: Record<string, true> = {
 	EXTERNAL_WRITE: true,
 	IMPORTED: true,
 };
-const SCRATCHPAD_MAX_ITEMS = Number.parseInt(process.env.PROMETHEUS_MEMORY_SP_MAX ?? "1000", 10);
+const SCRATCHPAD_MAX_ITEMS = Number.parseInt(process.env.MNEMOPROMETHEUS_SP_MAX ?? "1000", 10);
 
 function metadataJson(metadata: Metadata | null | undefined): string | null {
 	return metadata == null ? null : JSON.stringify(metadata);
@@ -191,7 +193,7 @@ function proactiveLinkIfEnabled(
 	content: string,
 	extractEntities: boolean,
 ): void {
-	if (process.env.PROMETHEUS_MEMORY_PROACTIVE_LINKING !== "1") return;
+	if (process.env.MNEMOPROMETHEUS_PROACTIVE_LINKING !== "1") return;
 	try {
 		const graph =
 			beam.episodicGraph instanceof EpisodicGraph
@@ -248,6 +250,96 @@ function rowToDict(row: Row): Row {
 	return { ...row };
 }
 
+/** Re-embedding batch size for a model-change rebuild — bounds each background
+ *  embedding request instead of embedding the whole corpus in one call. */
+const EMBED_REBUILD_BATCH = 128;
+
+/**
+ * Reconcile stored embeddings against the active embedding model at store open.
+ *
+ * Every `memory_embeddings` row is stamped with the model that produced it (see
+ * `runEmbedding` in `helpers.ts`). When the configured embedding model changes,
+ * its vector dimension changes too, so the previously-stored vectors are no
+ * longer comparable. On a mismatch we wipe every stored vector — the
+ * `memory_embeddings` table, the `episodic_memory.binary_vector` column, and the
+ * sqlite-vec `vec_episodes` index — then enqueue all live memories for
+ * background re-embedding under the new model via `scheduleEmbedding`.
+ *
+ * Runs once per store open; a fresh store (no embeddings) or an already-current
+ * store is a no-op. The destructive wipe is skipped whenever it could not be
+ * rebuilt — embeddings disabled via the runtime option OR the
+ * `MNEMOPROMETHEUS_NO_EMBEDDINGS` env, or an unresolved (empty) active model — so a
+ * stale-but-valid corpus is never destroyed without a replacement. MUST run
+ * inside the active runtime-options scope so `currentEmbeddingModel()` /
+ * `embeddingsDisabled()` reflect the per-instance configuration.
+ */
+export function reconcileEmbeddingModel(beam: BeamMemoryState): void {
+	if (embeddingsDisabled()) return;
+	const active = currentEmbeddingModel().trim();
+	if (active === "") return;
+
+	// Re-embed in bounded batches so a corpus-wide rebuild never issues one giant
+	// embedding request; each batch is its own tracked background task.
+	const rebuild = (items: readonly EmbedItem[]): void => {
+		for (let offset = 0; offset < items.length; offset += EMBED_REBUILD_BATCH) {
+			scheduleEmbedding(beam, items.slice(offset, offset + EMBED_REBUILD_BATCH));
+		}
+	};
+
+	// Stop at the first row whose stamped model differs from the active one
+	// (NULL/unstamped counts as a mismatch via `IS NOT`).
+	const mismatch = beam.db.query("SELECT 1 FROM memory_embeddings WHERE model IS NOT ? LIMIT 1").get(active);
+	if (mismatch) {
+		const staleModels = beam.db
+			.query("SELECT DISTINCT model FROM memory_embeddings WHERE model IS NOT ?")
+			.all(active) as { model: string | null }[];
+		const live = beam.db
+			.query(`
+				SELECT id AS memoryId, content FROM working_memory WHERE superseded_by IS NULL
+				UNION ALL
+				SELECT id AS memoryId, content FROM episodic_memory WHERE superseded_by IS NULL
+			`)
+			.all() as EmbedItem[];
+
+		transaction(beam.db, () => {
+			beam.db.prepare("DELETE FROM memory_embeddings").run();
+			beam.db.prepare("UPDATE episodic_memory SET binary_vector = NULL").run();
+			if (vecAvailable(beam.db)) {
+				try {
+					beam.db.prepare("DELETE FROM vec_episodes").run();
+				} catch {
+					// sqlite-vec cleanup is best-effort; rebuild correctness takes precedence.
+				}
+			}
+		});
+
+		logger.info("mnemopi: embedding model changed, rebuilding", {
+			from: staleModels.map(row => row.model ?? "(unstamped)"),
+			to: active,
+			count: live.length,
+		});
+		rebuild(live);
+		return;
+	}
+
+	// No stale embeddings, but a previously-interrupted rebuild (a failed embed or a process
+	// exit after the wipe) can leave live memories with no active-model embedding. Treating an
+	// empty/partial table as "reconciled" would strand them FTS-only, so re-enqueue any live
+	// row still missing an active-model embedding.
+	const missing = beam.db
+		.query(`
+			SELECT id AS memoryId, content FROM working_memory
+			WHERE superseded_by IS NULL AND id NOT IN (SELECT memory_id FROM memory_embeddings WHERE model = ?)
+			UNION ALL
+			SELECT id AS memoryId, content FROM episodic_memory
+			WHERE superseded_by IS NULL AND id NOT IN (SELECT memory_id FROM memory_embeddings WHERE model = ?)
+		`)
+		.all(active, active) as EmbedItem[];
+	if (missing.length === 0) return;
+	logger.info("mnemopi: resuming interrupted embedding rebuild", { to: active, count: missing.length });
+	rebuild(missing);
+}
+
 export function remember(beam: BeamMemoryState, content: string, options: StoreRememberOptions = {}): string {
 	const source = options.source ?? "conversation";
 	const importance = options.importance ?? 0.5;
@@ -594,7 +686,7 @@ export function scratchpadClear(beam: BeamMemoryState): void {
 export function exportToDict(beam: BeamMemoryState): Record<string, unknown> {
 	const db = beam.db;
 	return {
-		prometheus_memory_export: {
+		mnemopi_export: {
 			version: "1.0",
 			export_date: toUtcIso(),
 			source_db: beam.dbPath ?? ":memory:",

package/src/core/binary-vectors.ts

@@ -99,7 +99,7 @@ function isReadonlyMap(
 }
 
 export function getVecType(env: NodeJS.ProcessEnv = process.env): VecType {
-	const value = (env.PROMETHEUS_MEMORY_VEC_TYPE ?? "int8").trim().toLowerCase();
+	const value = (env.MNEMOPROMETHEUS_VEC_TYPE ?? "int8").trim().toLowerCase();
 	if (value === "float32" || value === "int8" || value === "bit") {
 		return value;
 	}

package/src/core/content-sanitizer.ts

@@ -19,9 +19,9 @@ export interface BlobMetadata {
 }
 
 export function blobRoot(env: NodeJS.ProcessEnv = process.env): string {
-	return env.PROMETHEUS_MEMORY_BLOB_DIR && env.PROMETHEUS_MEMORY_BLOB_DIR.length > 0
-		? env.PROMETHEUS_MEMORY_BLOB_DIR
-		: join(homedir(), ".prometheus", "memory", "blobs");
+	return env.MNEMOPROMETHEUS_BLOB_DIR && env.MNEMOPROMETHEUS_BLOB_DIR.length > 0
+		? env.MNEMOPROMETHEUS_BLOB_DIR
+		: join(homedir(), ".hermes", "mnemopi", "blobs");
 }
 
 export function computeSha256(data: Uint8Array | string): string {

package/src/core/cost-log.ts

@@ -3,7 +3,7 @@ import { mkdirSync } from "node:fs";
 import { homedir } from "node:os";
 import { dirname, join } from "node:path";
 
-export const DEFAULT_LOG_DIR = join(homedir(), ".prometheus", "memory", "data");
+export const DEFAULT_LOG_DIR = join(homedir(), ".mnemopi", "data");
 export const DEFAULT_LOG_DB = join(DEFAULT_LOG_DIR, "cost_log.db");
 
 export interface CostStats {

package/src/core/embeddings.ts

@@ -1,8 +1,9 @@
 import { mkdirSync } from "node:fs";
+import { type ApiKey, ProviderHttpError, withAuth } from "@prometheus-ai/ai";
+import { hostMatchesUrl } from "@prometheus-ai/catalog/hosts";
 import {
 	$env,
 	$flag,
-	APP_DISPLAY_NAME,
 	extractHttpStatusFromError,
 	fetchWithRetry,
 	getFastembedCacheDir,
@@ -11,7 +12,13 @@ import {
 import type { EmbeddingModel } from "fastembed";
 import { LRUCache } from "lru-cache/raw";
 import packageJson from "../../package.json" with { type: "json" };
-import { type EmbeddingOutput, getMnemopiRuntimeOptions, resolveEmbeddingProvider } from "./runtime-options";
+import { loadFastembed } from "./fastembed-runtime";
+import {
+	type EmbeddingOutput,
+	getMnemopiRuntimeOptions,
+	mnemopiDebugEnabled,
+	resolveEmbeddingProvider,
+} from "./runtime-options";
 
 export type { EmbeddingOutput } from "./runtime-options";
 export { cosineSimilarity } from "./vector-math";
@@ -55,17 +62,7 @@ const providerIds = new WeakMap<object, number>();
 let nextProviderId = 1;
 
 async function defaultLocalModelInitializer(options: LocalModelInitOptions): Promise<LocalEmbeddingModel> {
-	// Preload ORT 1.24 before fastembed's bundled ORT 1.21 — only on Windows,
-	// where loading the older binding first triggers a DLL-reuse crash. The 1.24
-	// line also has no darwin/x64 prebuilt, so importing it unconditionally breaks
-	// the darwin-x64 `bun build --compile` (Bun folds process.platform/arch and
-	// fails to resolve a binding that doesn't ship). The `win32` literal guard is
-	// statically foldable, so Bun dead-code-eliminates this import on every
-	// non-Windows target; fastembed loads its own ORT 1.21 binding there.
-	if (process.platform === "win32") {
-		await import("onnxruntime-node");
-	}
-	const { FlagEmbedding } = await import("fastembed");
+	const { FlagEmbedding } = await loadFastembed();
 	return FlagEmbedding.init(options);
 }
 
@@ -102,20 +99,30 @@ function inTestRuntime(): boolean {
 	return $env.NODE_ENV === "test" || $env.BUN_ENV === "test";
 }
 
-function embeddingsDisabled(): boolean {
+export function embeddingsDisabled(): boolean {
 	const active = activeEmbeddingOptions();
 	if (active?.disabled !== undefined) {
 		return active.disabled;
 	}
-	return $flag("PROMETHEUS_MEMORY_NO_EMBEDDINGS");
+	return $flag("MNEMOPROMETHEUS_NO_EMBEDDINGS");
 }
 
-function embeddingApiKey(): string {
+function embeddingApiKey(): ApiKey {
 	const active = activeEmbeddingOptions();
 	if (active?.apiKey !== undefined) {
 		return active.apiKey;
 	}
-	return $env.PROMETHEUS_MEMORY_EMBEDDING_API_KEY || $env.OPENROUTER_API_KEY || $env.OPENAI_API_KEY || "";
+	return (
+		$env.MNEMOPROMETHEUS_EMBEDDING_APROMETHEUS_KEY ||
+		$env.OPENROUTER_APROMETHEUS_KEY ||
+		$env.OPENAI_APROMETHEUS_KEY ||
+		""
+	);
+}
+
+/** A resolver always counts as configured; a static key only when non-empty. */
+function embeddingKeyConfigured(key: ApiKey = embeddingApiKey()): boolean {
+	return typeof key === "function" || key !== "";
 }
 
 function embeddingBaseUrl(): string {
@@ -123,7 +130,7 @@ function embeddingBaseUrl(): string {
 	if (active?.apiUrl !== undefined) {
 		return active.apiUrl;
 	}
-	return $env.PROMETHEUS_MEMORY_EMBEDDING_API_URL || $env.OPENROUTER_BASE_URL || "https://openrouter.ai/api/v1";
+	return $env.MNEMOPROMETHEUS_EMBEDDING_APROMETHEUS_URL || $env.OPENROUTER_BASE_URL || "https://openrouter.ai/api/v1";
 }
 
 function defaultModel(): string {
@@ -131,14 +138,14 @@ function defaultModel(): string {
 	if (active?.model !== undefined) {
 		return active.model;
 	}
-	return $env.PROMETHEUS_MEMORY_EMBEDDING_MODEL || "BAAI/bge-small-en-v1.5";
+	return $env.MNEMOPROMETHEUS_EMBEDDING_MODEL || "BAAI/bge-small-en-v1.5";
 }
 
 /**
  * Resolve the embedding model name for the currently active runtime scope.
  *
  * Reads (in order): the active provider's `model` from `withMnemopiRuntimeOptions`,
- * the `PROMETHEUS_MEMORY_EMBEDDING_MODEL` env var, then the bundled fastembed default. Stored
+ * the `MNEMOPROMETHEUS_EMBEDDING_MODEL` env var, then the bundled fastembed default. Stored
  * alongside each row in `memory_embeddings.model` so migrations can re-embed when
  * the active model changes.
  */
@@ -155,11 +162,11 @@ export function isApiModel(modelName: string): boolean {
 		return true;
 	}
 	const active = activeEmbeddingOptions();
-	const baseUrl = active?.apiUrl ?? ($env.PROMETHEUS_MEMORY_EMBEDDING_API_URL || $env.OPENROUTER_BASE_URL);
-	if (baseUrl !== undefined && baseUrl !== "" && !baseUrl.includes("openrouter.ai")) {
+	const baseUrl = active?.apiUrl ?? ($env.MNEMOPROMETHEUS_EMBEDDING_APROMETHEUS_URL || $env.OPENROUTER_BASE_URL);
+	if (baseUrl !== undefined && baseUrl !== "" && !hostMatchesUrl(baseUrl, "openrouter")) {
 		return true;
 	}
-	return $flag("PROMETHEUS_MEMORY_EMBEDDINGS_VIA_API");
+	return $flag("MNEMOPROMETHEUS_EMBEDDINGS_VIA_API");
 }
 
 const MODEL_DIMS: Record<string, number> = {
@@ -182,7 +189,7 @@ const MODEL_DIMS: Record<string, number> = {
 	"jina-embeddings-v5-omni-small": 1024,
 };
 export function embeddingDimFor(modelName: string): number {
-	const override = Number.parseInt($env.PROMETHEUS_MEMORY_EMBEDDING_DIM ?? "", 10);
+	const override = Number.parseInt($env.MNEMOPROMETHEUS_EMBEDDING_DIM ?? "", 10);
 	if (Number.isFinite(override)) {
 		return override;
 	}
@@ -239,7 +246,11 @@ async function getLocalModel(): Promise<LocalEmbeddingModel | null> {
 	localModelPromise = loading;
 	try {
 		return await loading;
-	} catch {
+	} catch (error) {
+		logger[mnemopiDebugEnabled() ? "warn" : "debug"]("mnemopi: local embedding model failed to load", {
+			model: modelName,
+			error: String(error),
+		});
 		if (localModelPromise === loading) localModelPromise = null;
 		return null;
 	}
@@ -247,31 +258,41 @@ async function getLocalModel(): Promise<LocalEmbeddingModel | null> {
 
 async function embedApi(texts: readonly string[]): Promise<EmbeddingMatrix | null> {
 	const baseUrl = embeddingBaseUrl();
-	const isCustom = !baseUrl.includes("openrouter.ai");
+	const isCustom = !hostMatchesUrl(baseUrl, "openrouter");
 	const apiKey = embeddingApiKey();
-	if (!isCustom && apiKey === "") {
+	if (!isCustom && !embeddingKeyConfigured(apiKey)) {
 		return null;
 	}
 
-	const headers: Record<string, string> = {
-		"Content-Type": "application/json",
-		"User-Agent": `${APP_DISPLAY_NAME}/${packageJson.version}`,
-		"HTTP-Referer": "https://prometheus.trivlab.com/",
-		"X-OpenRouter-Title": APP_DISPLAY_NAME,
-		"X-OpenRouter-Categories": "cli-agent",
-	};
-	if (apiKey !== "") {
-		headers.Authorization = `Bearer ${apiKey}`;
-	}
-
+	const body = JSON.stringify({ model: defaultModel(), input: texts });
 	try {
-		const response = await fetchWithRetry(`${baseUrl.replace(/\/+$/, "")}/embeddings`, {
-			method: "POST",
-			headers,
-			body: JSON.stringify({ model: defaultModel(), input: texts }),
-			signal: AbortSignal.timeout(30000),
-			maxAttempts: 3,
-			defaultDelayMs: attempt => 2 ** attempt * 1000,
+		// withAuth re-resolves the key on 401 (force-refresh, then sibling
+		// rotation) when `apiKey` is a resolver. The 429 backoff stays inside
+		// the attempt via fetchWithRetry. An empty static key attempts without
+		// an Authorization header (local/proxy setups).
+		const response = await withAuth(apiKey, async key => {
+			const headers: Record<string, string> = {
+				"Content-Type": "application/json",
+				"User-Agent": `Oh-My-Pi/${packageJson.version}`,
+				"HTTP-Referer": "https://prometheus.sh/",
+				"X-OpenRouter-Title": "Oh-My-Pi",
+				"X-OpenRouter-Categories": "cli-agent",
+			};
+			if (key !== "") {
+				headers.Authorization = `Bearer ${key}`;
+			}
+			const res = await fetchWithRetry(`${baseUrl.replace(/\/+$/, "")}/embeddings`, {
+				method: "POST",
+				headers,
+				body,
+				signal: AbortSignal.timeout(30000),
+				maxAttempts: 3,
+				defaultDelayMs: attempt => 2 ** attempt * 1000,
+			});
+			if (res.status === 401) {
+				throw new ProviderHttpError("mnemopi embedding request unauthorized (401)", 401, { headers: res.headers });
+			}
+			return res;
 		});
 		if (!response.ok) {
 			return null;
@@ -335,11 +356,11 @@ export async function available(): Promise<boolean> {
 		return providerAvailable(providerOverride);
 	}
 	if (isApiModel(defaultModel())) {
-		const baseUrl = active?.apiUrl ?? ($env.PROMETHEUS_MEMORY_EMBEDDING_API_URL || $env.OPENROUTER_BASE_URL);
-		if (baseUrl !== undefined && baseUrl !== "" && !baseUrl.includes("openrouter.ai")) {
+		const baseUrl = active?.apiUrl ?? ($env.MNEMOPROMETHEUS_EMBEDDING_APROMETHEUS_URL || $env.OPENROUTER_BASE_URL);
+		if (baseUrl !== undefined && baseUrl !== "" && !hostMatchesUrl(baseUrl, "openrouter")) {
 			return true;
 		}
-		return embeddingApiKey() !== "";
+		return embeddingKeyConfigured();
 	}
 	if (inTestRuntime()) {
 		return false;
@@ -348,7 +369,7 @@ export async function available(): Promise<boolean> {
 }
 
 export function availableApi(): boolean {
-	return embeddingApiKey() !== "";
+	return embeddingKeyConfigured();
 }
 
 export async function embedQuery(text: string): Promise<Vector | null> {
@@ -410,7 +431,11 @@ export async function embed(texts: readonly string[]): Promise<EmbeddingMatrix |
 			}
 		}
 		return vectors;
-	} catch {
+	} catch (error) {
+		logger[mnemopiDebugEnabled() ? "warn" : "debug"]("mnemopi: local embedding failed", {
+			textCount: texts.length,
+			error: String(error),
+		});
 		return null;
 	}
 }