Skip to content

armgit / qm

Public Imported from GitHub
0

Multiplayer agent harness for work.

Keep memory vectors across gateway moves and bound index writes (#2278)

Regan Bell104b32d88324
diff --git a/src/memory/semantic-recall.ts b/src/memory/semantic-recall.ts
index 896430839..61be9de63 100644
--- a/src/memory/semantic-recall.ts
+++ b/src/memory/semantic-recall.ts
@@ -5,6 +5,8 @@ import type { MemoryService, MemoryRecallContext } from "./memory-service.ts";
import { factCandidate, factHash, recentCandidates, type MemoryCandidate, type MemoryFact } from "./facts.ts";
import type { MemoryEmbedder } from "../model/embeddings.ts";
+const INDEX_WRITE_LANES = 4;
+
export interface MemoryVector {
scope: string;
model: string;
@@ -33,6 +35,10 @@ export function createSemanticMemoryService(
opts: { minSimilarity?: number; onError?: (error: unknown) => void } = {},
): MemoryService {
const queue = createKeyedQueue<string>();
+ const writeLanes = createKeyedQueue<number>();
+ let nextLane = 0;
+ const writeAll = <T>(items: readonly T[], write: (item: T) => Promise<void>): Promise<void[]> =>
+ Promise.all(items.map((item) => writeLanes(nextLane++ % INDEX_WRITE_LANES, () => write(item))));
const vectorId = ({ scope, model, hash }: Omit<MemoryVector, "vector">) => JSON.stringify([scope, model, hash]);
const cache = new Map<string, number[]>();
const remember = (model: string, hash: string, vector: number[]) => {
@@ -59,7 +65,7 @@ export function createSemanticMemoryService(
remember(row.model, row.hash, row.vector);
}
}
- await Promise.all(stale.map((row) => index.delete(vectorId(row))));
+ await writeAll(stale, (row) => index.delete(vectorId(row)));
}
const missing = [...new Set(facts.map((fact) => fact.text))].filter((text) => !vectors[factHash(text)]);
for (let start = 0; start < missing.length; start += 64) {
@@ -75,7 +81,7 @@ export function createSemanticMemoryService(
hash: factHash(text),
vector: embedded[i]!,
}));
- await Promise.all(rows.map((row) => index.put(vectorId(row), row)));
+ await writeAll(rows, (row) => index.put(vectorId(row), row));
for (const row of rows) {
vectors[row.hash] = row.vector;
remember(row.model, row.hash, row.vector);
@@ -110,12 +116,10 @@ export function createSemanticMemoryService(
}
const queryVector = await request;
if (facts.some((fact) => vectors[factHash(fact.text)]?.length !== queryVector.length)) {
- await Promise.all(
- Object.keys(vectors).map((hash) => {
- cache.delete(`${embedder.id} ${hash}`);
- return index.delete(vectorId({ scope, model: embedder.id, hash }));
- }),
- );
+ await writeAll(Object.keys(vectors), (hash) => {
+ cache.delete(`${embedder.id} ${hash}`);
+ return index.delete(vectorId({ scope, model: embedder.id, hash }));
+ });
throw new Error("Memory embedding dimensions changed; rebuilding the index on the next recall");
}
return facts
diff --git a/src/model/embeddings.ts b/src/model/embeddings.ts
index 6b94e6af2..f96d92be8 100644
--- a/src/model/embeddings.ts
+++ b/src/model/embeddings.ts
@@ -36,7 +36,7 @@ export function createMemoryEmbedder(
const target = routed?.target ?? model.id;
return {
id: createHash("sha256")
- .update(JSON.stringify([url, target]))
+ .update(JSON.stringify(routed ? [model.provider, target] : [url, target]))
.digest("hex"),
async embed(texts, signal) {
const response = await fetchImpl(url, {
diff --git a/test/semantic-memory.test.ts b/test/semantic-memory.test.ts
index 598f577d7..43e6b596a 100644
--- a/test/semantic-memory.test.ts
+++ b/test/semantic-memory.test.ts
@@ -190,3 +190,75 @@ test("a provider vector-width change falls back to recency and rebuilds the inde
[3],
);
});
+
+test("moving the embedding gateway to another URL keeps the vector index", async (t) => {
+ const { base, index } = await fixture(t);
+ await base.capture("personal:a", ["browser preference", "release note"], AT);
+ const model = {
+ kind: "embedding" as const,
+ baseUrl: "https://gateway.invalid",
+ id: "gateway/embed",
+ name: "embed",
+ provider: "qm:gateway",
+ };
+ const embedded: string[] = [];
+ const through = (url: string) => {
+ const transport = { url, apiKey: "k", apiKeyHeader: "x-k", models: { "gateway/embed": "embed" } };
+ const embedder = createMemoryEmbedder(model, null, transport, async (_url, init) => {
+ const input = JSON.parse(String(init?.body)).input as string[];
+ embedded.push(...input);
+ return Response.json({
+ data: input.map((text, i) => ({ index: i, embedding: /browser|verify/i.test(text) ? [1, 0] : [0, 1] })),
+ });
+ });
+ return createSemanticMemoryService(base, async () => embedder, index);
+ };
+ assert.equal(
+ await recall(through("http://signer.one:8080/v1"), "personal:a", { query: "verify" }),
+ "- (2026-10-01) browser preference",
+ );
+ const rows = await index.all();
+ embedded.length = 0;
+ assert.equal(
+ await recall(through("http://signer.two:8080/v1"), "personal:a", { query: "verify" }),
+ "- (2026-10-01) browser preference",
+ );
+ assert.deepEqual(embedded, ["verify"]);
+ assert.deepEqual(await index.all(), rows);
+});
+
+test("re-indexing a large notebook for a new model keeps index writes bounded", async (t) => {
+ const { base, index } = await fixture(t);
+ await base.capture(
+ "personal:a",
+ Array.from({ length: 40 }, (_, i) => `fact ${i}`),
+ AT,
+ );
+ let inFlight = 0;
+ let peak = 0;
+ const track = async (write: () => Promise<void>) => {
+ peak = Math.max(peak, ++inFlight);
+ await new Promise((resolve) => setImmediate(resolve));
+ try {
+ await write();
+ } finally {
+ inFlight--;
+ }
+ };
+ const tracked = {
+ ...index,
+ put: (id: string, value: MemoryVector) => track(() => index.put(id, value)),
+ delete: (id: string) => track(() => index.delete(id)),
+ };
+ const model = (id: string): MemoryEmbedder => ({ id, embed: async (texts) => texts.map(() => [1, 0]) });
+ for (const id of ["old-model", "new-model"])
+ await recall(
+ createSemanticMemoryService(base, async () => model(id), tracked),
+ "personal:a",
+ { query: "fact" },
+ );
+ assert.ok(peak <= 4, `peak concurrent index writes ${peak}`);
+ const rows = await index.all();
+ assert.equal(rows.length, 40);
+ assert.ok(rows.every((row) => row.model === "new-model"));
+});