Avinash-jetwani/jevmem: Automatic mission reminiscence for Claude Code. Also works with Cursor and Codex. · GitHub

Automatic mission reminiscence for Claude Code. Also works with Cursor and Codex.

• Saves selections, constraints, bugs and todos out of your Claude Code chats into JEVMEM.md, mechanically.
• When you modify your thoughts, the outdated line is marked outmoded, not deleted.
• Next session, the related traces are added to Claude’s context.

- [decision] Use Postgres 16 for the first retailer; SQLite locks underneath load  
- [constraint] Node 20 is the ground; CI runs 20 and 22  
- [superseded] Use SQLite as the first retailer → id:k3d9xq  

npm set up -g jevmem
export TYPESAFE_API_KEY=...        # https://typesafe.ai (an OpenAI or Anthropic secret's non-compulsory)
cd your-project
jevmem init --tool claude

init creates JEVMEM.md, jevmem.config.json and a gitignored .jevmem/ folder, and registers two Claude Code hooks in .claude/settings.native.json, which it provides to .gitignore (details).

What is computerized and what depends upon the agent:

MCP add_memory goes by way of the identical gate because the hook. Client configs: docs/mcp.md.

1. Scrub. Common secret shapes, e-mail addresses and card-shaped numbers are faraway from the flip earlier than it leaves your machine.
2. Ask Jev typed questions. Jev by TypeSafe AI solutions a hard and fast set of small questions with chances: is there a choice, a rule, a bug? is it small discuss or an injection try? which current line does it change?
3. Apply thresholds in code. Plain guidelines over these chances determine save or skip; they stay in jevmem.config.json, not in a immediate.
4. Write one line. On save, a small LLM (or a deterministic extract, with no LLM key) writes one line of at most 200 characters.
5. Supersede the outdated line. If the flip replaces an current reminiscence, that line is tagged [superseded] … → id:new and stays within the file.

Tiers, questions, coverage, contradictions, recall and audit: docs/how-it-works.md.

66 held-out turns, all seven deciders given the identical state, 2026-09-23 (method, regression set, pricing, p95, retries):

The 0.30 s is the Jev API determination; by way of an actual Stop hook course of, Node start-up included, it’s 0.6 s finish to finish (cost and latency).

On 66 held-out turns, jevmem’s median determination took 0.30 s, in opposition to 2.8–4.3 s for six present LLMs.
Its accuracy was throughout the LLMs’ vary: 98.5% save/skip (tied with GPT-6 Astra for highest) and 95.5% save+form, in opposition to 90.9–98.5% for the LLMs. GPT-6 Astra (98.5%) and Claude Opus 5.5 (97.0%) have been extra correct on save+form; Claude Fable 5.1 tied; GPT-6 Luna, Gemini 3.8 Flash and Grok 4.7 have been much less correct. It discovered 5/5 contradictions, as did 5 of the six LLMs.
GPT-6 Luna was cheaper ($0.000089 in opposition to $0.000127) however much less correct (93.9%) and about 10× slower.
This is a single run, and variations of 1 or two turns are inside run-to-run noise. If essentially the most correct determination issues most, GPT-6 Astra or Claude Opus 5.5 are higher, at about 40–60× the fee per determination and 9–12× the latency. jevmem is for whenever you need a quick, low-cost determination on each message.

• Sent to TypeSafe AI: the person message of every flip (and the assistant reply for questions and bug stories), the earlier two turns, and your reminiscence traces, to be scored. No telemetry. If you set an OpenAI or Anthropic key, the textual content of a saved flip additionally goes to that supplier to put in writing the road.
• Scrubbed first: widespread credential shapes (API keys, tokens, *_PASSWORD= type pairs, connection-string passwords, personal keys), e-mail addresses and 16-digit numbers; names, telephone numbers and addresses are usually not caught.
• Zero-retention flag: jevmem can ship zeroDataRetention: true (computerized for Vercel AI Gateway URLs); whether or not it applies depends upon the gateway and TypeSafe’s phrases, and jevmem doesn’t confirm it.

Exactly what is shipped, saved and scrubbed: SECURITY.md.

• Early: v0.4; each eval units have been written by the creator, and neither is an impartial benchmark.
• Not essentially the most correct: GPT-6 Astra and Claude Opus 5.5 scored larger on save+form; jevmem’s edge is velocity and value.
• Recall high quality just isn’t measured: that related traces are injected is examined; whether or not solutions get higher just isn’t.
• Long-run drift just isn’t measured: the harness covers five-turn periods, not weeks of use.
• Automatic seize is Claude Code solely (and Codex whereas jevmem watch runs); Cursor and Claude Desktop save solely when the agent calls add_memory.
• Jev outages drop turns: every Jev name has a 2 s price range; when the API is gradual or down, the flip is skipped and logged in .jevmem/log.jsonl, not retried later.

jevmem init [--tool claude|cursor|codex|claude-desktop|all] [--no-hooks] [--command ""]
jevmem hook                                    Hook entrypoint; reads the Claude Code hook JSON on stdin
jevmem daemon [status|start|stop]              Warm Jev shopper utilized by the hook (auto-started, exits when idle)
jevmem watch [--replay] [--once]               Capture turns from Codex's session log for this mission
jevmem mcp [--root ]                      Stdio MCP server
jevmem audit [--dry-run]                       Re-score each reminiscence in opposition to the repo, flag [stale?]
jevmem search  [--limit N]              Rank recollections by relevance
jevmem record [--all]                            Print recollections
jevmem add                         Add a line by hand (secrets and techniques scrubbed; no Jev examine)
jevmem why                            Every Jev reply behind a line or a skipped flip
jevmem proper                          Label a choice as right
jevmem unsuitable  [--should-be ]   Label a choice as unsuitable
jevmem missed "" [--kind ]         Label a flip that ought to have been saved
jevmem match [--dry-run] [--force]               Refit weights and thresholds from labels (wants 40+)
jevmem stats                                   Latency p50/p95, price per day, cache hit charge, escalation charge, labels, final match
jevmem log                                     Per-label latency, token and value abstract of .jevmem/log.jsonl

Every command accepts --help. Set JEVMEM_VERBOSE=1 for a one-line latency/price abstract after each hook run.