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

Jevmem is a CLI tool that automatically extracts and persists important project-level memories - decisions, constraints, bugs and todos - directly from Claude Code chats (with optional support for Cursor and Codex). It initializes a JEVMEM.md file plus config and a gitignored .jevmem/ folder, registers hooks into Claude Code, and writes compact, single-line memory entries (≤200 chars). When a memory is superseded it is not deleted but marked as superseded with a reference to the new line. Install and run via npm, provide a Typesafe API key (OpenAI/Anthropic optional), and use commands to init, audit, search, add, label, and inspect Jev decisions. Automatic capture differs by tool: Claude Code captures every turn, Codex can capture while jevmem watch runs, and Cursor/desktop setups rely on agent-initiated calls.

Decisions use a local pipeline: scrub sensitive shapes, send limited context to Jev (TypeSafe AI) which answers typed questions, apply deterministic thresholds in config, and either skip or write one line; saved turns optionally pass through an LLM to render the line. A benchmark on 66 held-out turns shows a 0.30 s median decision time (0.6 s end-to-end), 98.5% save/skip accuracy and 95.5% save+kind, at roughly $0.000127 per decision - far cheaper and faster than larger LLMs, though not the absolute most accurate. Privacy notes, scrubbing rules, and honest limits (v0.4 early-stage, internal evals, unmeasured long-run recall drift, outages dropping turns) are documented.