Jetwani Avinash (@jetwaniavinash) on X

How it works

After each turn, jevmem asks Jev, a model from TypeSafe AI, whether the turn contained anything worth keeping: a rule, a decision, something that failed and why, or a to-do. If yes, it appends one line to a plain Markdown file in your project. If not, it writes nothing. The page's example rules are "Dates are day-month-year" and "Don't push without asking."

Two lifecycle details that the summary does not cover:

• Retiring a line. You can retire a line you no longer need. It stays in the file, so git history is intact, but Claude stops receiving it. Retiring is the mechanism for pruning memory without deleting it.
• Guarding commands. Before Claude runs a command, jevmem checks it against saved rules (for example "never push a tag without asking"). By default it asks you first. Blocking is opt-in per rule: you "trust a rule" and it then blocks instead of asking.

Because the memory is plain Markdown, you can also edit lines by hand and review every change in git.

Evidence

All figures come from the author's own tests. The page gives no description of the setup.

• Rule adherence: across 72 "real sessions", Claude followed the rule in 66 with jevmem. Without it the count was 28. The page does not say whether those 28 are out of the same 72 sessions.
• Save/skip decisions: Jev made the right save-or-skip call on 98.5% of turns in a 66-turn test. That is roughly one wrong call.
• Command guard: of 274 test commands, 68 broke a saved rule. jevmem caught 66 of those 68, so 2 rule-breaking commands got through.

Limits

• The page does not report false positives for the guard: how often it flagged any of the 206 compliant commands. That number determines how noisy the ask-first default will be.
• All tests are small and self-run. The page does not say which rules, projects or agents were used, or whether Cursor and Codex were tested at all rather than only Claude Code.
• The save/skip decision depends on an external model from a third party (TypeSafe AI). The page does not say where that model runs, what the plugin sends to it, or what it costs.
• The author explicitly asks users to report "where it saves the wrong thing", which treats wrong saves as a known, open failure mode.
• The page gives no install command. The only link is a shortened t.co URL.