# Brainfeather

> Long-term memory for AI coding agents. Brainfeather records the durable facts about your project — stack, conventions, decisions — and hands them back to your agent on its next run, so each session does not start from zero.

Brainfeather is a memory layer that sits beneath AI coding assistants such as
Claude Code, Cursor, opencode and Antigravity, and connects over the Model
Context Protocol (MCP). A language model's context window is RAM: fast, and
wiped when the session closes. Brainfeather is the disk beside it.

## Status

In early development, and free while it is. Access is invite-only: request one
from the waitlist on the home page. Approved testers create an account, generate an API key,
and connect it to an MCP-compatible editor. The published package is
@brainfeather/mcp on npm. Pin npx -y @brainfeather/mcp@1.5.1, then run
npx -y @brainfeather/mcp@1.5.1 init so recall is automatic. Hosted Streamable
HTTP MCP is at https://brainfeather.com/mcp.

## What it does

- Captures durable project facts in the background, off the critical path, so the session never waits on extraction.
- Client → MCP → dashboard: inferred facts wait at /review until approved; they do not enter recall on their own.
- Retires stale facts rather than merely outranking them: change a decision and the previous one is marked invalid, so later reads return only what still holds.
- Filters conversational noise. Greetings and thinking-out-loud are not stored; tech choices, project rules and conventions are.
- Shares one store across every connected client, so a fact written from one tool is readable by the others on their next run.

## What it is not

- Not a general-purpose personalisation or user-profile platform. The focus is software projects and the agents working on them.
- Not a vector database. It is a memory layer with opinions about what is worth keeping.
- Not a replacement for your agent's context window — it is the durable store beside it.

## Pages

- [Home](https://brainfeather.com/): what it is and how it works.
- [Request access](https://brainfeather.com/#waitlist): join the invite-only waitlist.
- [Hosted MCP](https://brainfeather.com/mcp): Streamable HTTP for clients that cannot run local stdio.
- [API keys](https://brainfeather.com/api-keys): create a key, then pin @brainfeather/mcp@1.5.1 and run init.
- [Review queue](https://brainfeather.com/review): inferred captures wait here until approved. They do not enter recall on their own.
- [Contact](https://brainfeather.com/contact): support, security reports, press.
- [Privacy Policy](https://brainfeather.com/privacy): what is collected and stored, and for how long.
- [Terms of Service](https://brainfeather.com/terms): the terms of use.

## Contact

getbrainfeather@gmail.com