@memloom/core

The engine behind memloom, a local-first memory engine for AI agents. Embed it in your own application, or use the memloom CLI instead if you want the daemon, viewer and MCP server ready to go.

npm install @memloom/core

Usage

import { Memloom, PgliteAdapter } from "@memloom/core";

const storage = await PgliteAdapter.open({ dataDir: "./data" });
const memloom = new Memloom({ storage, embedding, llm });
await memloom.init();

await memloom.save({ content: "the staging database runs on Postgres" });
const hits = await memloom.recall("staging db");

init() runs the migrations and verifies the store's embedding fingerprint, so a store embedded with one model refuses to open under another.

What is in it

• The belief pipeline. Saves dedupe, restatements become versions, and contradictions are kept and raised rather than overwritten.
• Hybrid retrieval. Vector, keyword and entity-graph arms fused by rank in a single SQL function, over memories and ingested documents together.
• Context ingestion. Markdown, text, PDF, web pages, audio and video, chunked along real structure with citations. The extractor registry is open, so you can add a format.
• The entity graph. Schema-constrained extraction, reversible entity folds, traversal.
• Storage adapters. PgliteAdapter for embedded Postgres in a folder, PgAdapter for a real server with pgvector. Identical schema and SQL on both.

Core reads no environment variables and holds no global state. Every provider and setting is a constructor argument, so the whole engine is testable without a network.

Documentation

docs.memloom.dev/concepts/architecture

Apache-2.0. Built by Versuno.