Jevmem – automatic project memory for Claude Code, built on Jev | AI Paper Digest

TL;DR Highlight

Saving chat decisions to JEVMEM.md raised Claude's adherence from 28/72 to 66/72 sessions.

Who Should Read

Developers whose Claude Code sessions forget last week's decisions and keep re-explaining them.

Core Mechanics

Chat decisions saved as one-line repo notes
- jevmem writes decisions, rules, bugs, to-dos and dead ends to JEVMEM.md as single lines.
- Only lines relevant to your prompt are sent back to Claude next session.
- Teammates share the same file through git.
Jev scores what to save with probabilities
- Jev, a TypeSafe AI model, answers yes/no questions like 'is this a decision or a rule?' with probabilities.
- Thresholds in jevmem.config.json decide save or skip in code, not a prompt.
- Secrets, emails and card-like numbers are scrubbed before sending.
Following past decisions: 28/72 to 66/72
- Across 72 real Claude Code sessions, Claude followed earlier decisions in 28 without memory and 66 with jevmem.
- Forbidden changes attempted dropped from 10 of 18 to 0 of 18.
- Repeating failed approaches dropped from 3 of 15 to 0 of 15.
A hand-kept CLAUDE.md did about as well
- The same lines in a hand-written CLAUDE.md scored 67 of 72 versus jevmem's 66.

Community Reactions

Opinion that the purpose is unclear
- One commenter read the project and still couldn't tell what problem it solves or why it beats the harness's native memory.
Pushback on keeping superseded lines
- A commenter said marking old lines superseded instead of deleting them pollutes context with outdated info.
- They said pruning conflicting info is the biggest weakness of built-in memory.
Privacy concern about TypeSafe AI
- A commenter saw a lot of data going to TypeSafe AI in SECURITY.md and found it a red flag.
Simpler alternative and KV cache question
- One commenter suggested a cheap agent with one JSON-schema-constrained tool call instead.
- Another asked whether it thrashes the session's KV cache.

How to Apply

Preview your CLAUDE.md with import
- Run jevmem import in a current project to see what it would add from CLAUDE.md.
- Add --apply to write it, and read PRIVACY.md first to see what is sent.
Write failed approaches as one-line rules
- Where the agent repeats mistakes, add an 'tried, failed, reason' line to CLAUDE.md or JEVMEM.md.
- Check next session whether Claude still repeats the approach.
Test the guard with a forbidden rule
- Save a rule like 'Never commit .env files' and ask for a related command.
- Check whether a warning appears and note wordings that slip through.

Try It

copy & run

# Paste into Claude Code (or Claude/ChatGPT) — simplified version of jevmem's approach (manual project memory extraction)

Below is this chat's history. Extract only items worth reusing next session in JEVMEM.md format.
Each item is one line, at most 200 characters.
Prefix each with one of [decision], [constraint], [bug], [todo], [dead-end].
For failed approaches, include the reason.
If a message reverses an earlier decision, mark the old line [superseded].

{{chat history}}

Next step: save the result to JEVMEM.md in your repo and start the next session with "Read JEVMEM.md first and follow its rules".

Terminology

JevA TypeSafe AI model that answers yes/no questions with probabilities. It acts like a judge scoring text rather than generating it.

dead endAn approach that was tried and failed, with the reason. It is a 'road closed' sign so nobody goes down it again.

always-on ruleA rule every change must follow, sent to Claude at the start of every session regardless of the prompt.

MCPA standard way for an AI tool to call another program. Think of it as a USB port for AI tools.

SessionStart hookA hook that runs when a Claude Code session starts. jevmem uses it to inject always-on rules.

Jevmem – automatic project memory for Claude Code, built on Jev