Memhub

memhub

*Your AI agents forget things. memhub doesn't.*

Every new session, your AI agent starts from zero. You re-explain the build command, the naming conventions, and why you chose Postgres over SQLite six months ago. memhub gives the agent a persistent, searchable memory of your project, so it looks those things up instead of asking you again.

memhub is one offline binary and one SQLite file next to your code. Claude Code, Codex, and OpenCode all share it. It holds decisions (with the *why* behind them), facts, tasks, your own reference docs, and an index of where things live in your codebase. When the agent needs context, it pulls a small ranked bundle of just the relevant rows, not your whole README. Prompts stay short and your token bill stays down. There is no cloud, no account, and no daemon, and everything, including semantic search, runs locally.

Jump to Quickstart → · See how it works →

Plain-word definitions of the technical terms used below

- MCP (Model Context Protocol): a standard way for an agent to call a local tool, such as memhub, as a structured function.
- Skills: the short instruction files, such as /wrap-up and /recall, that memhub installs for each agent.
- FTS5: SQLite's built-in full-text search. It finds rows by the words they contain.
- BM25: the formula FTS5 uses to rank rows by how well their words match a query.
- BGE-small: the small embedding model bundled in the binary. It turns text into a list of numbers, so rows with similar meaning end up close together.
- Hybrid recall: a recall that blends keyword search (FTS5 with BM25) and meaning-based search (BGE-small), then re-ranks the best candidates. The alternative, FTS-only recall, uses keyword search alone.
- Re-ranker: a second small model that reads the query together with one…