Breaking

Jevmem Releases Open-Source Memory Layer for Claude Code, Cursor and Codex

October 7, 2026 at 15:46 EDT

AI Agents
Software Dev & Coding
Open Source

Jevmem, an MIT-licensed project-memory tool for coding agents, has been released publicly after being promoted as a framework built with 1.4 billion Claude Opus 5.5 tokens of development work. The project writes durable repo-level memories into JEVMEM.md so agents can carry decisions, constraints and failed approaches across sessions.

Oct. 6, 2026 · jevmem

A memory file for coding agents

Jevmem saves decisions, rules, bugs and failed approaches into JEVMEM.md, then brings relevant lines back when Claude Code, Cursor or Codex needs them.

1.4B

Opus 5.5 tokens cited in the launch claim

66/72

Sessions followed the saved decision

66/68

Rule breaks caught by the guard

0.25s

Median save-or-skip decision time

Memory changes the outcome

Decision-following sessions out of 72, shown as proportional stacked columns.

How the loop works

Decide

A turn may contain a decision, rule or dead end.

Score

Jev assigns typed probabilities.

Write

Jevmem updates JEVMEM.md.

Recall

Relevant lines return in a later prompt.

Why developers care

It targets repeated decisions, forbidden edits and failed approaches that coding agents often forget between sessions.

What remains unproven

The tests are author-run, answer quality is not measured, and the guard is a backstop rather than a sandbox.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.