Jevmem

Say it once. jevmem saves the decisions, rules and failed approaches from your Claude Code chats to JEVMEM.md in your repo, and brings the relevant ones back next session. Before Claude runs a command or edits a file, it checks the call against your saved rules. It is open source (MIT), built on TypeSafe AI's Jev, in Anthropic's Claude plugin directory and on the MCP Registry. Since 0.7.4 it saves from Cursor chats the same way, and it works with Codex and Claude Desktop over MCP.

In Anthropic's Claude plugin directory · On the MCP Registry · Open source, MIT

The docs site, one page per question: avinash-jetwani.github.io/jevmem

https://github.com/user-attachments/assets/c8a7eced-c427-4ef6-85f2-124eb34ad291

Why

Each Claude Code session starts with a fresh context, so what you decided last week lives in last week's chat. A CLAUDE.md file helps if you keep it up to date. jevmem keeps a file like it up to date for you, as you work.

In numbers

In real Claude Code sessions on three small test projects, with jevmem and without it:

• Claude followed the project's earlier decisions: in 66 of 72 sessions with jevmem, 28 of 72 without (about 9 in 10, against 4 in 10).
• Tried a change the project forbids: in 0 of 18 sessions with jevmem, against 10 of 18 without.
• Repeated an approach that had already failed: in 0 of 15, against 3 of 15.
• Saving: about 18 cents per 1,000 messages and about a fifth of a second to decide, in the background: Claude doesn't wait for it.
• Bringing lines back: about a third to half a second on each prompt, which Claude waits for; the time and the cost grow with the file.
• To be fair: a CLAUDE.md you keep up to date by hand did about as well (67 of 72), and better on one convention nothing in the prompt points at (3 of 3 against 0 of 3). jevmem keeps a file like it up to date for you, and since 0.7.1 marks such rules always-on, so they reach every session: 18 of 18 on six held-out tasks written for that gap, against 14 of 18 with 0.7.0.

My own test sets, each number from one run; none is an independent benchmark. Method and results

How it works

You chat as usual. When you decide something, jevmem writes it as one line in a file in your repo, and brings the right lines back next session.

1. You decide something in a chat: "We use Postgres."
2. jevmem asks Jev by TypeSafe AI, a model that answers yes/no questions with probabilities, whether it's worth keeping.
3. If it is, jevmem writes one line to JEVMEM.md in your repo.
4. Next session, the lines that matter for your prompt go back to Claude.

Change your mind, and the old line is crossed out: kept for history, not sent to Claude. Your team gets the same file through git.

jevmem keeps decisions, rules, bugs, to-dos and dead ends (an approach that was tried and failed, with the reason).

- [superseded] We'll use SQLite as the primary store for now. → id:cuasaq
- [decision] Switch the primary store to Postgres 16.
- [constraint] Node 20 is the minimum supported version, and CI runs Node 20 and 22.

Each line also carries a comment with its id, time and confidence, left out above. The full lines:

- [superseded] We'll use SQLite as the primary store for now. → id:cuasaq  
- [decision] Switch the primary store to Postgres 16.  
- [constraint] Node 20 is the minimum supported version, and CI runs Node 20 and 22.  

Real lines from 0.5.7's default writer, which 0.5.8 to 0.5.10 did not change (the run, 2026-09-26). It kept one sentence of each turn: the Postgres turn also said "SQLite locks up under concurrent writes", and that reason was left out. Since 0.6.0 the line is made from the sentences Jev picks, the one that states the memory and the one that gives its reason (What's new). With writer set, an OpenAI or Anthropic model condenses the whole turn instead.

Does it work?

In 72 real Claude Code sessions for each setup, Claude acted on the saved line about as often with jevmem as with the same lines in a hand-written CLAUDE.md, and in fewer than half the sessions with neither.

So jevmem did about as well as a hand-written CLAUDE.md, without you writing it. CLAUDE.md did better on a convention nothing in the prompt points at (3 of 3 against 0 of 3), measured on 0.6, where jevmem brought a line back only for a prompt it bears on. Since 0.7.1 jevmem marks a rule every change must follow always-on and sends it at the start of every session: on six held-out tasks written for that gap, Claude followed the convention in 14 of 18 sessions with 0.7.0 and 18 of 18 with 0.7.1, as with the same lines in a hand-kept CLAUDE.md (what's new, limits). How the sessions were run, and every row: the benchmark page.

The guard

Before Claude runs a command or edits a file, jevmem checks it against your saved rules. If one might break, Claude Code asks you first:

jevmem: this may break a saved rule: "Never commit .env files" (JEVMEM.md)

On a test of 274 tool calls, a set the guard was not tuned on, it caught 66 of 68 rule breaks and asked about 3 or 4 of the 206 calls that broke no rule (one run each on 0.6.0 and 0.6.1). It's a backstop, not a sandbox: it looks at the words a rule and a call share. Since 0.7.0, jevmem trust lets a rule you added by hand block a call in the guard's block mode, not only ask about it. How the guard works.

Side pane

A pane beside your Claude Code chat shows what jevmem just did: the memory lines it gave Claude for your prompt, what its guard said about a command, and the line it saved from the turn, each with why, wrong and forget buttons. The two that change JEVMEM.md ask first; a rule is retired in a terminal only. It reads three local files and sends nothing anywhere.

/plugin install jevmem-pane@jevmem

It is a second plugin from the jevmem marketplace (the jevmem plugin itself still ships no code), for Claude Code 2.1.299 or newer; it opens on its own in a terminal 144 columns wide or wider and in the desktop app's Code tab, and /jevmem opens or closes it anywhere. What each block shows, and what was measured: docs/pane.md.

Install

You need a TypeSafe API key, which jevmem's requests are billed to, and Claude Code 2.1.273 or later.

1. In the Claude app: Plugins → Discover → jevmem → Add.
2. Install the CLI the plugin runs:

   npm install -g jevmem
   

3. In your project's folder, finish the setup in one go (it asks for your key without showing it, turns jevmem on in that folder, and checks the setup):

   jevmem setup
   

Start Claude Code in that project. If you start it before the setup is done, the plugin says once what is left; jevmem doctor checks the setup any time.

From the jevmem marketplace:

npm install -g jevmem
claude plugin marketplace add Avinash-jetwani/jevmem
claude plugin install jevmem@jevmem
claude plugin install jevmem-pane@jevmem   # the side pane (0.7.5), optional
cd your-project && jevmem setup

With npm only (also sets up Cursor and Codex):

npm install -g jevmem
cd your-project
jevmem init --tool claude    # or cursor, codex, claude-desktop, all

Already have a CLAUDE.md? jevmem import shows what it would add from it; --apply writes it.

Every step, and what to do if the plugin can't find the CLI: docs/install.md.

Works with

Saving is automatic in Claude Code and, since 0.7.4, in Cursor, where Cursor's hooks save each finished agent turn; and in Codex while jevmem watch runs. Cursor, Codex and Claude Desktop reach jevmem over MCP, a standard way for an AI tool to call another program, when the agent asks for it; in the Cursor app a project's MCP server starts off until you turn it on once (the step). In Cursor, on 10 held-out chats captured before the code (24 turns), the first run paired 21 of 24 turns whole, the prompt with the whole reply; a reader fix came from that run, and the second run paired 24 of 24, with what to save right in 21 of 24 (measured).

The MCP server is on the MCP Registry as io.github.Avinash-jetwani/jevmem. Client setup: docs/mcp.md · docs/install.md.

Fast and cheap

Deciding what to save takes 0.23 s and costs $0.00018 per message, in the background: Claude doesn't wait for it. On whether to save a message at all, jevmem was right as often as the best of six LLMs (98.5%); two of them were better at naming the kind of line.

Measured on 66 turns written before the code and not tuned on, one run of 0.7.5 on 2026-10-17 (results); the six LLMs ran on 2026-09-23. All seven in one table, with the cost of each and how it was run: benchmark.

How it decides

No prompt decides what to save: Jev answers small yes/no questions with probabilities, and plain rules in code act on them.

1. Scrub. Common secret shapes, email addresses and card-shaped numbers are removed from the turn before it leaves your machine.
2. Ask Jev typed questions. Jev by TypeSafe AI answers a fixed set of small questions with probabilities: is there a decision, a rule, a bug? is it small talk or an injection attempt? does it say the same as a saved line (0.7.0)? is it a rule for how any work in this project is done (0.7.1)? which existing line does it change?
3. Apply thresholds in code. Plain rules over those probabilities decide save or skip; they live in jevmem.config.json, not in a prompt. A message that restates a saved line is skipped as a duplicate, and one that restates it word for word is skipped before anything is sent (0.7.0).
4. Write one line. On save, jevmem writes one line of at most 200 characters from the sentences of the turn that Jev picks (the one that states the memory and the one that gives its reason), or, if you set writer in jevmem.config.json, a small OpenAI or Anthropic model condenses the turn.
5. Supersede the old line. If the turn replaces an existing memory, that line is tagged [superseded] … → id:new and stays in the file.
6. Mark it always-on (0.7.1). A saved rule that Jev reads as one every change must follow gets always:1 in its comment, and the SessionStart hook puts it in front of Claude at the start of every session, whatever the prompt. jevmem always <id> marks a line by hand.

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

Privacy

Your messages go to TypeSafe's API to be scored, with common secrets removed first, and nowhere else unless you set a writer.

• jevmem only runs in projects you turn on (jevmem enable or jevmem init). Elsewhere, nothing is sent.
• Your message, the previous two turns and your memory lines go to TypeSafe's API to be scored, with common secrets scrubbed first. So does Claude's reply on turns that ask a question, report a bug or try something, and, for the guard, the command or the file being changed.
• jevmem import --from sessions (0.7.1) sends a project's past Claude Code sessions, turn by turn and scrubbed the same way, only when you run it and answer yes after seeing the count and the cost.
• No telemetry. Nothing goes to OpenAI or Anthropic unless you set writer in jevmem.config.json.
• Lines a teammate or a pull request adds are checked before Claude sees them.

Exactly what is sent, scrubbed and checked, with the third parties' privacy policies and how to delete your data: PRIVACY.md. What the check on planted lines does not cover: SECURITY.md.

Limits

jevmem is young, and these are the limits to know before you rely on it:

• Early: first released on 2026-09-22, and every test set was written by the author or captured from Claude Code sessions on scratch projects built for it. None is an independent benchmark.
• Not the most accurate: two LLMs scored higher at picking the kind of line. jevmem's edge is speed and cost.
• Answer quality isn't measured: the tests check that the right lines reach Claude, not that its answers get better.
• Automatic saving is Claude Code and, since 0.7.4, Cursor (and Codex while jevmem watch runs); Claude Desktop saves only when the agent calls add_memory, and in Cursor recall is still the agent's call.
• The guard looks at shared words: a rule worded far from the command it should catch can be missed.
• The side pane is Claude Code only (2.1.299 or newer): Cursor has no pane, and Claude Code for VS Code draws none yet.

Every limit, with the numbers: docs/limits.md.

What's new

What changed, newest first; the measurements behind each line are on the what's new page.

• 0.7.5: a pane beside the chat shows what jevmem did. Install jevmem-pane from the jevmem marketplace and Claude Code shows, for each prompt and turn, the lines it recalled, what the guard decided and what was saved, with why, wrong and forget buttons. Also: a leading label like "Decision for this project:" no longer ends up in the saved line (on 12 held-out turns written that way, run once, 0 of 10 lines kept it), jevmem init --tool cursor gitignores Cursor's own state folder, jevmem doctor says when the Cursor app has not started the project's MCP server, and jevmem wrong … --should-be none on a rule asks on a terminal like forget does (what's new).
• 0.7.4: a Cursor chat saves like a Claude Code one. jevmem init --tool cursor writes Cursor's hooks (and a launcher, both committable), and when an agent turn finishes the stop hook reads the turn from Cursor's transcript and runs the same decide and write path as in Claude Code; reading memory in Cursor stays the agent's MCP call. On 10 held-out Cursor chats (24 turns), captured through Cursor's own CLI before the code, the first run paired 21 of 24 turns whole; a reader fix came from that run, and the second run paired 24 of 24, with what to save right in 21 of 24. Run jevmem init --tool cursor once more in a project set up earlier. Details.
• 0.7.3: the setup message stays off cloud and Cowork sessions. 0.7.2's one-time message showed at the start of every claude.ai chat that ran in a cloud workspace, where its marker could not stay. Now it shows only on your own computer (a terminal, the desktop app, the VS Code extension or an SSH session), and a cloud or Cowork session gets no message and no file. The plugin's launcher and the docs; the CLI's code is unchanged. Details.
• 0.7.2: finish the setup in one command. After Add in the Claude app, the first session on a machine says once what is left, and jevmem setup asks for your key, turns jevmem on in the folder and checks the setup. Always-on lines now go through the poisoning gate before they reach a session, and Jev 1.13.0 is pinned: beside 1.13 on six test sets, 1.14 missed one of five contradictions in all three runs, saved fewer dead ends from captured sessions in all three, and kept fewer of the reasons on the written set. Details.

Details and measurements: docs/whats-new.md · Upgrading: docs/upgrading.md · CHANGELOG

jevmem init [--tool claude|cursor|codex|claude-desktop|all] [--no-hooks] [--command "<cmd>"]
jevmem init --remove-hooks                     Remove jevmem's Claude Code hooks and Cursor hooks from this project (plugin users)
jevmem enable                                  Opt this project in (plugin users): jevmem.config.json, JEVMEM.md, .jevmem/
jevmem setup                                   Finish the setup here: the CLI and the plugin, your TypeSafe key (asked for, not shown), jevmem turned on in this folder, doctor; terminal only
jevmem disable                                 Opt this project out: jevmem does nothing here (JEVMEM.md is kept)
jevmem hook                                    Hook entrypoint; reads the Claude Code hook JSON on stdin
jevmem daemon [status|start|stop]              Warm Jev client used by the hook (auto-started, exits when idle)
jevmem watch [--replay] [--once]               Capture turns from Codex's session log for this project (Cursor saves through its hooks)
jevmem mcp [--root <dir>]                      Stdio MCP server
jevmem audit [--dry-run]                       Re-score every memory against the repo, flag [stale?]
jevmem audit --security [--ci]                 List lines that read as instructions to an AI (--ci: exit 1 if any)
jevmem search <query> [--limit N]              Rank memories by relevance
jevmem list [--all]                            Print memories (--all: with superseded and retired lines, and provenance)
jevmem add [--trust] [--always] <kind> <text>  Add a line by hand (secrets scrubbed, a leading tag dropped; no Jev check); --trust marks it verified (asks on a terminal); --always sends it to Claude in every session
jevmem always [--off] <id> [<id>…]             Mark a line always-on (in Claude's context at the start of every session), or take the mark off
jevmem always --suggest [--apply]              Ask Jev which live lines are rules for every change and propose the marks; --apply sets them
jevmem forget <id> [<id>…]                     Retire a line in place: it stays in JEVMEM.md as [retired]; a rule asks for a yes on a terminal
jevmem trust <id> [<id>…]                      Mark a line you wrote as verified, after the poisoning gate (asks on a terminal); the guard can then block on it
jevmem import [--from <sources>] [--apply]     Import CLAUDE.md, AGENTS.md, .cursor/rules/* (claude-auto-memory on request); dry run by default
jevmem import --from sessions [--since 30d] [--apply]   Read this project's past Claude Code sessions; counts and asks before anything is sent
jevmem why <id|hash>                           Every Jev answer behind a line or a skipped turn
jevmem right <id|hash>                         Label a decision as correct
jevmem wrong <id|hash> [--should-be <kind|none>]   Label a decision as wrong (--should-be none retires the line)
jevmem missed "<text>" [--kind <kind>]         Label a turn that should have been saved
jevmem fit [--dry-run] [--force]               Refit weights and thresholds from labels (needs 40+)
jevmem stats                                   Writer, latency p50/p95, cost per day, cache hit rate, escalation rate, retry queue, labels, last fit
jevmem doctor                                  Is this project enabled, where the TypeSafe key comes from, which writer is active and why
jevmem key                                     Save your TypeSafe API key to ~/.jevmem/env (asks for it without showing it)
jevmem log                                     Per-label latency, token and cost summary of .jevmem/log.jsonl
jevmem guard test "<command>" | --edit <path>  Dry run of the PreToolUse guard on one call: rules, prefilter, Jev's answer, hook output
jevmem guard log [-n 20]                       The guard's recent asks and denials in this project, with the rule and score

These are 0.7.3's commands: 0.5.10's, jevmem guard (0.6.0), forget, trust and add --trust (0.7.0), always, add --always and import --from sessions (0.7.1), and setup (0.7.2); 0.7.3 added no command and changed none. Every command accepts --help. Set JEVMEM_VERBOSE=1 for a one-line latency/cost summary after every hook run.

Links

• Docs: how it works · install · the guard · dead ends · benchmark · cost · hooks · MCP and client configs · configuration · limits · demo
• CHANGELOG · Releases · DECISIONS · CONTRIBUTING · SECURITY · PRIVACY
• Built on Jev by TypeSafe AI. License: MIT