Four calls, two shapes of context and a tag index — enough to hand an assistant a project's history and get it back next session, with entries that can name what they continue. The author says so on the front page: it never went past v0.1.0, and Claude Code's native memory files and CLAUDE.md now cover the same ground. Read it if you are building a context store of your own; look elsewhere if you want one that is still being maintained.
A stdio server that gives an assistant somewhere to put context between sessions. Two shapes go in: project context, which carries a project id and can point at a parent context and at related entries, and conversation context, which carries a session id and can mark itself the continuation of an earlier one. Both are tagged, stored as JSON on your own disk, and read back by id or listed by project, tag or kind. Its author archived the project in March 2026, naming Claude Code's own memory files and CLAUDE.md as having taken the job over.
- Project context saved under a project id, with an optional parent context and a list of related context ids
- Conversation context saved under a session id, optionally marked as the continuation of a previous one
- Free-form tags and a metadata object on every entry
- Any entry retrieved by id, scoped to its project when you have one
- Entries listed and filtered by project, by tag, or by kind — project or conversation
Node, and the repository itself: the compiled server is committed alongside the source, so it runs from the checkout. No account and no key — contexts are JSON files on your own disk, under a directory the server creates. The project is archived and stopped at v0.1.0, so what is in the repository is the finished shape.