A memory tool for Claude Code claims sub-second, cent-level decisions

jevmem is a new open-source tool that automatically saves decisions, constraints, bugs and todos from Claude Code sessions into a JEVMEM.md file, then re-injects the relevant lines into the next session's context. Rather than asking a large model to judge what is worth remembering, jevmem routes each turn through a small classifier service, called Jev, that scores it in about 300 milliseconds, and applies fixed thresholds written in code rather than a prompt. In the author's own benchmark of 66 held-out turns, jevmem matched GPT-6 Astra's 98.5 percent save-or-skip accuracy at roughly 300 milliseconds and $0.000127 per decision, against 2.8 to 4.3 seconds and $0.005 to $0.013 per decision for six current models asked to make the same call directly. On the finer save-plus-kind judgment, jevmem trailed GPT-6 Astra and Claude Opus 5.5 by a few points.

The author states plainly that this is not an independent result: "both eval sets were written by the author, and neither is an independent benchmark," and the tool is labeled v0.4 and "early." Using it means sending each turn's text, the prior two turns, and existing memory lines to Jev's API for scoring, after stripping common secret and email patterns first; a saved memory line is then optionally written using an OpenAI or Anthropic key supplied by the user.

Why it matters: The bet here, that a fast, cheap, threshold-based classifier can do nearly as well as routing every turn through a frontier model, is a concrete answer to a cost and latency problem teams running agents at scale actually have. It comes with a tradeoff its own author states clearly: chat text leaves the machine to be scored by a third party, and the benchmark backing the claim is self-graded rather than independently verified.