Maker here. I got tired of re-explaining my stack to every AI tool, so I built one memory they all share: tell Claude Code once, Cursor already knows. What's different under the hood: Itsuki doesn't replay transcripts — it extracts typed objects, source-links each one to the message it came from, and versions everything so you can roll a fact back. Relationships need verbatim evidence: if the model can't quote the sentence naming both endpoints, the edge is rejected with a named reason. The model proposes; the backend decides. Honest limits: hosted (Cloudflare), not local-first; export covers one memory space; writes are async; no third-party security audit yet. Engine is Apache 2.0 — happy to answer anything, including where the design is wrong.

Itsuki

The memory layer under every AI tool you use.

The memory layer under every AI tool you use.