A persistent memory primitive that follows you across Claude.ai, ChatGPT, the CLI, and your tools.
Project Memory, ChatGPT Memory, Claude Memory. Three vendors converging on the same lock-in. The thing each one calls “memory” lives inside their product and only inside their product. Switch vendors and it does not come with you.
Brain is the inversion. One server. Your Postgres. A single MCP endpoint that every Anthropic and OpenAI surface mounts. The data lives where you put it and follows you wherever you go.
Anyone running a serious LLM workflow accumulates context. Decisions. Project notes. Architectural patterns. That context becomes infrastructure, but the labs treat it as a feature. Brain treats it as infrastructure.
Tokenized recall with tag-weighted scoring. OR-match across content, tags, and category. A score prefix on every result so you can judge ranking quality. Streamable HTTP transport so a single Cloud Run instance can serve every client without sticky sessions.
A NestJS server with a clean Memory module sits in front of a Postgres database. The same module is wrapped twice. Once as a REST API for the CLI and GUI clients. Once as an MCP tool surface for Claude.ai, ChatGPT, and Claude Code custom connectors.
The MCP layer uses Streamable HTTP in stateless mode. The session ID generator is undefined. Each request brings its own auth and spins up a fresh transport. That means no session affinity is required on Cloud Run, so instances can scale to zero between requests.
The first version did a single LIKE on the whole query. “60 day financial health review” returned nothing. The rewrite tokenizes on word boundaries, OR-matches each token across content, tags, and category, and scores by match count with tag and category hits weighted higher. The score prefix on every result lets callers see ranking quality.
Persistent SSE connections combined with session affinity pinned instances permanently. That blocked Cloud Run from scaling down. Stateless Streamable HTTP serves every request fresh, lets instances die between calls, and eliminates the session-affinity requirement.
Earlier SSE work showed that the SDK-generated session ID is itself the auth token for follow-up POSTs. Applying middleware to both endpoints breaks claude.ai connectors because the SSE endpoint URL does not carry forward the original credentials. Auth on the handshake, sessionId on the follow-up.