claude-mem — как работает персистентная память

★ 8.0 · learning

how-it-works is a Claude Code skill that explains the mechanics of claude-mem: how observations are captured, when memory injection activates, and where all data lives. Every Read, Edit, and Bash operation Claude performs becomes a compressed observation; at session end these are summarized and auto-injected into future prompts, so the next session starts with context from the previous one — no re-explaining the codebase, no re-discovering past decisions. Memory injection kicks in from the second session onward in a given project; running /learn-codebase lets you front-load the entire repo into memory in a single pass (roughly 5 minutes, optional). Everything stays local under ~/.claude-mem — SQLite database, vector index, logs, and settings — with no data leaving the machine except calls to the configured AI provider (Claude, OpenRouter, or Gemini) used for compression.