Jevmem – automatic project memory for Claude Code, built on Jev

Jevmem is an automatic project memory system for Claude Code that functions with Cursor and Codex as well. Upon initialization, it generates JEVMEM.md, jevmem.config.json, and a .jevmem folder, as well as adding two Claude Code hooks to .claude/settings.local.json. This configuration is then added to .gitignore. The system relies on the MCP add_memory command, which undergoes the same gate as the hooks. Client configurations can be found in docs/mcp.md.

The system is evaluated on held-out turns, with seven deciders given the same state for consistency. The decision process takes 0.30 seconds through the Jev API, with an end-to-end time of 0.6 seconds, including Node startup. This is significantly faster than current LLMs, which take between 2.8 and 4.3 seconds. On 66 held-out turns, Jevmem's median decision time was 0.30 seconds, compared to the LLMs' range of 2.8-4.3 seconds.

Accuracy-wise, Jevmem performs on par with current LLMs, with a save/skip rate of 98.5%, tied with GPT-6 Astra for the highest. Its save+kind accuracy is 95.5%, matching GPT-6 Astra and Claude Opus 5.5, and surpassing other models. GPT-6 Astra and Claude Opus 5.5 were more accurate on the save+kind metric, while Claude Fable 5.1 and GPT-6 Luna tied. GPT-6 Luna, Gemini 3.8 Flash, and Grok 4.7 were less accurate.

In terms of cost, GPT-6 Luna is the cheapest at $0.000089 per decision, while Jevmem costs $0.000127. However, Jevmem is significantly faster, with about 40-60 times the speed and 9-12 times the latency of GPT-6 Luna. If the most accurate decision is paramount, GPT-6 Astra or Claude Opus 5.5 are preferable, but at a much higher cost and latency. Jevmem excels when a fast, cost-effective decision on each message is desired.