Quickstart

The fishmem package is the low-level TypeScript engine. It is different from @fishmem/sdk, which calls a FishMem service over HTTP.

Install

npm install fishmem

The embedded engine is TypeScript-first. Python applications use the first-party HTTP SDK described at /sdk/python.

Create the engine

The engine accepts explicit graph, vector, LLM, and embedder implementations. This example uses deterministic in-memory components so it can run without provider keys:

import { Memory } from "fishmem";

const memory = await Memory.create({
  graphStore: { provider: "memory" },
  vectorStore: { provider: "memory" },
  embedder: { provider: "mock" },
  llm: { provider: "mock" },
});

For persistent deployments, replace the in-memory stores with SQLite, Postgres, D1, Qdrant, Vectorize, or another supported adapter. See Configuration.

Store a memory

await memory.add("Alex prefers dark mode and writes TypeScript.", {
  userId: "alex",
  infer: false,
  metadata: { source: "onboarding" },
});

infer: false stores the submitted string verbatim and makes zero LLM calls. For conversational input, omit infer (or set it to true) to make one extraction call and store only refined canonical records.

Recall it

const { results } = await memory.search("How should I present examples to Alex?", {
  userId: "alex",
  limit: 5,
});

Search with the same scope used on write. Depending on configuration, recall can combine keyword, vector, temporal, and graph signals.

Use the service instead

If your application should call a shared or remotely deployed FishMem instance, use @fishmem/sdk. It provides the same client across Node, Bun, Deno, Cloudflare Workers, and Vercel.