@galdor/memory-sqlite

An embedded SQLite long-term memory store for galdor. Implements the core memory.Store interface, so it drops in behind a Retriever like the bundled InMemoryStore — but persists chunks in a local SQLite file with FTS5 lexical (BM25) search and brute-force cosine vector search.

Zero external services: it runs on the runtime's built-in SQLite (bun:sqlite on Bun, node:sqlite on Node ≥ 22.5).

Install

bun add @galdor/memory-sqlite   # or: npm install @galdor/memory-sqlite

Usage

import { openSqlite } from "@galdor/memory-sqlite";

const store = openSqlite("memory.db"); // or ":memory:" for a transient store

await store.add([
  { id: "c1", documentId: "d1", index: 0, text: "Quito is the capital of Ecuador.", embedding: vec, metadata: { lang: "en" } },
]);

// Vector search
const byVector = await store.retrieve({ embedding: queryVec, k: 5, filter: { lang: "en" } });
// Lexical (FTS5 BM25) search
const byText = await store.retrieve({ text: "capital of Ecuador", k: 5 });

await store.delete("d1"); // removes every chunk of document d1
store.close();

Compose it with an embedder via Retriever for text-only queries backed by vector search:

import { Retriever } from "@galdor/core/memory";
const retriever = new Retriever({ store, embedder, defaultK: 5 });
const hits = await retriever.retrieve({ text: "capital of Ecuador" });

Behavior

• add upserts by chunk id (idempotent), storing documentId/index/text
  • metadata. Rejects empty ids and vector-dimension mismatches.
• retrieve supports either a text query (FTS5 BM25) or an embedding query (cosine over stored vectors); filter is an exact-match AND over metadata.
• delete removes every chunk whose documentId matches.

License

Apache-2.0