jevmem

Checked, not fully reviewed. We ran its own toolchain and read the data path; we did not drive a live Claude Code session against it.

jevmem watches a coding-agent conversation and writes decisions, constraints, bugs and todos into a JEVMEM.md beside your project, then feeds the relevant lines back at the start of the next session. When you change your mind, the old line is marked [superseded] with a pointer rather than removed — which is the right shape for a file a human also reads.

With its own toolchain — pnpm, a build first, then the suite — it is 248 tests passing and one skipped, and tsc --noEmit is clean. (Our first attempt used npm and no build; five test files failed on spawnSync pnpm ENOENT and a missing dist. That was our environment, not the project: worth saying because a red suite is a claim about a project and this one did not earn it.)

The part worth crediting is what leaves the machine. The state is assembled rather than dumped: the user’s message always, the assistant’s reply only when the user asked a question or said nothing at all, at most 1,500 characters of previous turns, and a shortlist of existing memories chosen by word overlap rather than the whole file. Every piece goes through a secret scrubber, and the memory texts are scrubbed again before they are sent. Two tiers of questions run — nine broad ones every turn, thirty atomic ones only when the first tier is unsure — so the common turn is one request.

MIT, on npm, CI green, with a DECISIONS.md recording its own design calls.

For the same job with a different boundary, see Jev-Mem; for what a coding agent sends when the decision is about a tool call rather than a memory, dsh-jev-interceptor.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 26, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.