Question 1

Isn't this just CLAUDE.md with extra steps?

Accepted Answer

Close, and the README admits where CLAUDE.md wins. On the author's own A/B run, a hand-written CLAUDE.md followed the project's decisions 67 of 72 times against 66 for jevmem, and it did better on a convention nothing in the prompt pointed at, so rules every task must follow still belong in that file.

Question 2

Are those 66-of-72 numbers independent evidence?

Accepted Answer

No. Every eval set was written by the author, and the project's own table shows two models beating it on save-plus-kind while it tied the best on save-or-skip. The comparison is worth reading as the author's measurement, not as an independent benchmark of project memory.

Question 3

Will the guard block my commands from running?

Accepted Answer

Not reliably, and it is a backstop rather than a sandbox. On a held-out set of 274 tool calls it caught 66 of 68 rule breaks but produced 3 to 4 false asks across 206 fine calls, and it works by looking at the words a rule and a call share. A held command still waits on you, not on a hard denial.

Question 4

Does it work in Codex and Cursor the same way?

Accepted Answer

Not the same way. Automatic capture every turn and automatic recall every prompt are Claude Code behaviours; for Codex, Cursor, and Claude Desktop the content is saved and returned over MCP when the agent calls it or asks. Claude Code 2.1.273 or later plus a TypeSafe key are required either way.

Question 5

What actually gets written to JEVMEM.md?

Accepted Answer

One line per memory, at most 200 characters, tagged with its kind: decision, rule, constraint, bug, to-do, or dead end, an approach that was tried and failed with the reason. Each line carries an id, a timestamp, and a confidence in a trailing comment. A replaced line stays in the file marked superseded, kept for history and never sent back to the agent.

Question 6

What does the memory cost, and does it slow my session down?

Accepted Answer

Saving runs in the background and Claude does not wait for it: about 17 cents per 1,000 messages and roughly a quarter second to decide, with the Stop hook exiting in 12 to 14 ms and the Jev decision recorded about 0.28 s after it starts. Recall is the part you wait on, about a third to half a second per prompt, and both the time and the cost grow with the size of the file.

Question 7

How do I install it, and can it import an existing CLAUDE.md?

Accepted Answer

Add the jevmem plugin from the Claude app, install the CLI with npm install -g jevmem, save your key with jevmem key, then run jevmem enable in the project; jevmem doctor checks the setup. jevmem init --tool claude, cursor, codex, claude-desktop, or all does the wiring from npm alone, and jevmem import shows what it would add from an existing CLAUDE.md, with --apply to write it.

Jevmem

About this project

Highlights

Questions about this entry

Reactions & coverage

Related terms

More Jev resources like this

From the Jev community