Jevmem Adds Automatic Memory for Claude Code, Cursor and Codex

Jevmem adds automatic project memory to Claude Code. It saves decisions, constraints, bugs and todos from chat sessions into JEVMEM.md and loads relevant lines back into context next session. The same system also works with Cursor and Codex, using one memory file shared across all three tools, according to the GitHub repository.

Installation is via npm install -g jevmem. Running jevmem init --tool claude creates JEVMEM.md, jevmem.config.json and a gitignored .jevmem/ folder in the project. The repository is documented at Avinash-jetwani/jevmem.

For Claude Code, init registers two hooks, or automatic triggers, in .claude/settings.local.json. Capture runs every turn via the Stop hook. Recall runs every prompt via UserPromptSubmit. No manual export is required.

Updates preserve history. When a decision changes, Jevmem marks the old line as superseded rather than deleting it. JEVMEM.md lives in the user's repo so the whole team gets the same decisions.

The broader context here is familiar to anyone running coding agents across sessions. Default behavior is stateless, with no carryover from one thread to the next. Useful choices disappear, and teams end up re-litigating stack decisions, re-discovering edge-case bugs, or keeping hand-edited instruction files that drift from what was actually decided.

Looking at what this means for workflow, the design choices are the interesting part. Checking JEVMEM.md into version control while keeping .jevmem/ gitignored separates shared knowledge from local scratch state. That avoids polluting diffs, or lists of changes, with machine bookkeeping while keeping decisions reviewable in pull requests. Marking superseded lines preserves an audit trail, which matters when a constraint changes twice and no one remembers the original rationale.

In my view, the risk to watch is the same as with any automatic recall layer: precision and noise. Injecting relevant lines every prompt reduces repetition, but relevance filtering has to stay tight as JEVMEM.md grows. A shared file across Claude Code, Cursor and Codex helps consistency, yet one noisy entry propagates everywhere. Worth flagging for early teams is the need for periodic curation, even if capture is hands-off. If that balance holds, the payoff is less re-prompting and more continuity session to session.