jevmory — your coding agent's memory, with receipts

Local, zero-dependency memory layer for coding agents. Every remembered fact is a verbatim quote graded by TypeSafe Jev's calibrated confidence — and every number has a receipt.

jevmory audit MEMORY.md

your memory has 1 stale line, 1 wrong line, 1 unsupported line; 2 keep

LINE  VERDICT      CONF  SUPP  CONTRA  CLAIM
   5  STALE        0.82  0.34    0.71  The build runs on Bun; bun run build is the…
   6  WRONG        0.93  0.07    0.93  The test suite runs with pytest.
   7  UNSUPPORTED  0.85  0.06    0.05  Failed API requests retry up to five times…

receipts: run 1 · 1 api calls · 838 tokens · evidence: 0 facts, 6 statements

The problem

Your coding agent forgets everything at session end. The usual fix — a hand-maintained MEMORY.md — rots silently: stale lines, lines nobody ever said, lines the project outgrew, and no way to tell which is which. jevmory closes both ends: it remembers what was actually said (hooks ingest every session locally), and it audits the memory file against that evidence, verdict by verdict.

What it does — four pillars

• 🎯 Verbatim facts, zero generation — no LLM generation anywhere. Jev judges; code selects and composes. Facts are quotes with verbatim context, so the memory can't hallucinate.
• 🔌 Agent-agnostic — a one-command Claude Code plugin, an MCP server (status / recall / fact) for Codex CLI, Cursor, and any MCP client, and Codex notify hooks. Same store, same receipts.
• 🛡️ Local by architecture — Python 3 stdlib only, per-project SQLite, secrets redacted at rest before storage, and grading gated behind an explicit per-project opt-in. No opt-in → nothing leaves.
• 📄 Receipts for everything — raw judgments and usage are stored (runs/judgments tables), so any number in any report traces to the API call that produced it.

Quickstart

Requires Python ≥3.10 and nothing else — stdlib only, no pip deps.

Plug-and-play with Claude Code (no pip install — the plugin is self-contained; the repo is both marketplace and plugin):

/plugin marketplace add romiluz13/jevmory
/plugin install jevmory@jevmory

That one install gives you: a SessionEnd hook that ingests every session locally, a SessionStart hook that injects this project's top remembered facts into context (fresh sessions, resumes, and post-compaction restarts alike), the three read-only MCP memory tools, the /jevmory:memory command, and a jevmory CLI on the Bash tool PATH — all vendored from the plugin's own copy of the package. Remove it with /plugin uninstall jevmory.

Any MCP agent (Codex CLI, Cursor, …) — add the stdio memory server:

jevmory mcp        # or: python3 -m jevmory.mcp

Point your agent at it as a stdio MCP server (command: python3, args: ["-m", "jevmory.mcp"]). Three read-only, fully local tools; no API key, no egress — grading stays in the CLI where the opt-in gate lives.

From source:

git clone https://github.com/romiluz13/jevmory && cd jevmory

# in your project:
jevmory init                     # setup; the store appears on first ingest
jevmory install --agent claude   # prints the Claude Code SessionEnd hook JSON
jevmory install --agent codex    # prints the Codex notify snippet (--yes patches config)

# stay fully local (nothing ever leaves):
jevmory ingest --scan            # finds this project's transcripts and ingests
jevmory status                   # queued candidates, last ingest, errors
jevmory recall                   # the top facts, as the hook injects them

# or opt in to grading (needs $TYPESAFE_API_KEY):
jevmory init --enable-grading    # per-project opt-in marker
jevmory distill                  # grade queued candidates
jevmory audit MEMORY.md          # receipts for every memory line
jevmory resolve <id>             # answer a contradiction question

# closed network? run grading against a local server (no key):
jevmory audit MEMORY.md --backend kev   # jaredpalmer/kev on localhost

Add jevmory.md to your project's .gitignore if you don't want agent memory in version control — it's yours, not the repo's.

Try the demo first — offline, deterministic, zero cost:

python3 demo/run_demo.py

A planted-error fixture: demo/MEMORY.md has 5 memory lines — 3 with planted errors contradicted by the evidence in demo/transcript.jsonl. Judgments are pinned offline, so the errors are guaranteed present and the demo costs nothing. The live jevmory audit produces exactly the output shown above, with real Jev judgments behind the numbers.

How it works

flowchart LR
    CC["Claude Code<br/>(SessionEnd hook)"] -->|redacted at rest| DB[("per-project SQLite<br/>+ FTS")]
    CX["Codex CLI<br/>(notify hook)"] -->|redacted at rest| DB
    DB -->|"distill: opt-in Jev grading"| J["raw judgments<br/>(receipts)"]
    J --> F["facts:<br/>verbatim quotes"]
    F --> MD["jevmory.md<br/>(project root)"]
    MD -->|audit| R["STALE / WRONG /<br/>UNSUPPORTED / KEEP"]
    DB -->|"SessionStart recall"| CTX["agent context"]
    DB -->|"MCP: status · recall · fact"| MCP["any MCP client"]

Loading

Hooks ingest locally and always exit 0 — they can never break a session. distill extracts and dedupes sentence candidates, then (only with your opt-in) grades them: durability, category, significance, support, contradiction. audit anchors lines that match a stored fact verbatim as VERIFIED deterministically (zero API spend) and grades the rest against the store's evidence. jevmory.md lands at the project root — grouped by category, confidence-ordered, sentinel-guarded against silent overwrites.

Design laws and domain model: docs/DOMAIN.md.

Configuration & privacy

• Fully local mode — just never run init --enable-grading. Everything else (ingest, status, scan, recall, MCP) is local by architecture. No marker → candidates queue; jevmory status says so.
• What leaves, and what never does — the exact egress surface, the redaction list, and the honest caveat about verbatim quotes: docs/PRIVACY.md.
• Closed networks — --backend kev grades against a local wire-compatible server ($JEVMORY_KEV_ENDPOINT, default 127.0.0.1:8009); no key, no egress, opt-in still required.
• Benchmarks — budget-capped live smoke and the planted-truth fidelity benchmark: docs/benchmarks/live-smoke.md.

Status

Active. 595 offline tests (the same python -m unittest discover CI runs on 3.10–3.13) — no network, no API key, FakeJev including adversarial mode. Modules M0–M7 complete; pipeline live-verified against the real Jev API (probe, capped distill, audit). v0.2: two-stage audit with deterministic VERIFIED anchoring, vintage receipts, and the kev local grading backend. v0.3: plug-and-play — a self-contained Claude Code plugin (marketplace + hooks + MCP tools + command), an agent-agnostic MCP server for Codex/Cursor, and SessionStart recall injection; both plugin manifests pass claude plugin validate --strict.

See CONTRIBUTING.md to get involved, and CHANGELOG.md for history.

License

MIT — see LICENSE.