Jevmory

No generation step writes facts; Jev judges while code selects and composes verbatim quotes with context. That keeps stored memory from hallucinating wording the session never contained.

Install the self-contained Claude Code plugin for hooks, recall injection, MCP tools, and CLI, or point any MCP client at python3 -m jevmory.mcp for read-only status, recall, and fact tools. Hooks always exit 0 so they cannot break sessions.

Ingest, status, scan, recall, and MCP stay fully local with redaction before storage. Only distill and audit send graded content out, gated by init --enable-grading, with --backend kev available for closed networks.

Run the offline deterministic demo with pinned judgments to see the planted-error report at zero cost. Live distill and audit then store raw judgments and usage in runs and judgments tables so every number traces to its call.

Grading is opt-in per project with init --enable-grading and needs a TYPESAFE_API_KEY, or the kev local backend on 127.0.0.1:8009 for a closed network. Lines that match an active stored statement are labelled MATCHED at zero API spend, and every graded number plus its usage is stored in the runs and judgments tables as a receipt.

Python 3.10 or newer with the standard library only, no pip dependencies, MIT-licensed. The project is active with 606 offline tests run on 3.10 through 3.13 with no network or API key, a FakeJev including adversarial mode, and modules M0 through M7 complete, live-verified against the real Jev API.

Yes. A line matching an active stored statement is labelled MATCHED for zero spend, but the README says matching does not check current code or refresh a verification date. Provenance proves the quote was really said, not that it is still true, so treat the receipt as evidence of the source rather than of correctness.