Supermemory is a memory and context engine for AI software: a service that keeps what a user has told your application across sessions and hands back a compact version of it whenever you ask. It is meant for developers building assistants, agents, chat products or internal tools that currently start every conversation from nothing. The project is TypeScript, MIT licensed, has been going since February 2024, and has collected around 30,000 stars and 2,600 forks. You can call it as a hosted API or run the whole thing yourself.
What it does
The problem it addresses is simple to state: a model forgets everything the moment the chat window closes. Supermemory sits beside your application as the layer that remembers.
In practice that means three things. It extracts facts from what passes through it, rather than storing raw transcripts verbatim. It keeps a per-user profile that is updated as new facts arrive. And it drops information that has gone stale, so an old preference does not keep resurfacing after the user has changed their mind.
Retrieval is deliberately cheap to use: one call returns the full profile for a user, which the project reports at roughly 50 milliseconds. The README claims first place on three public AI-memory benchmarks — LongMemEval, LoCoMo and ConvoMem — along with 95% Recall@15 at a 99.4% reduction in context. Those are the project's own published numbers; the repository links to its research write-up if you want to check the methodology before believing them.
There is also a plugin route. Install it and Claude Code carries your context from one session to the next, without you wiring up the API yourself.
How it works
The repository is a monorepo built around Cloudflare's platform: Workers for the compute, KV for fast key-value access, and Pages for the hosted front end. The web app is a Remix application built with Vite and Tailwind CSS. Durable storage runs on Postgres through Drizzle ORM. That combination is why the latency figures are framed the way they are — the read path is designed to be served from an edge runtime rather than a round trip to a single region.
Client-side, you do not talk to any of that directly. There is a supermemory package on npm and a supermemory package on PyPI, so a TypeScript or Python application adds memory as a dependency and a few calls. A console at console.supermemory.ai gives you the managed version with a dashboard.
The self-hosting story is the other half. The project advertises a single binary with zero configuration, and keeps a dedicated self-hosting section in its documentation. Running fully locally is presented as a first-class mode, not an afterthought, which matters when the data being remembered is personal.
Getting started
The documentation site carries a quickstart, and the fastest path is the managed API: create an account on the console, install the SDK for your language, and start writing memories and reading profiles. Nothing about your model provider has to change — this sits alongside whatever you already call.
If you would rather keep the data, follow the self-hosting overview instead and run the binary. The repository is the same code in both cases. A Discord community is linked from the README for questions, and the codebase is still receiving pushes as of September 2026.
When to use it / when not
Reach for it when your product's value grows with what it knows about a person: a coding assistant that should learn your conventions, a support agent that should not ask the same question twice, a companion app where continuity is the point. It is also a reasonable fit when your context windows are getting expensive and you would rather send a distilled profile than a growing pile of history.
It is less useful in a few situations. If your application is genuinely stateless — a one-shot classifier, a batch transform — there is nothing to remember and this is a dependency you do not need. If your retrieval problem is really document search over a fixed corpus rather than evolving facts about a user, a plain vector store is the simpler tool. And if you are running the hosted version, you are handing user facts to a third party, so the self-hosted path is worth the extra setup in regulated environments.
One practical caution: the benchmark claims come from the project itself. Test recall on your own data before you commit to it as the memory layer for something in production.
Anyone building a product where an AI is supposed to know its user should take this repository seriously. It picks a narrow, real problem, gives it a small API surface, offers both a managed and a fully local deployment, and is permissively licensed so trying it costs little more than an afternoon. The self-hosted option in particular makes it one of the few credible ways to add persistent memory without sending personal data anywhere you do not control.