bettermemory

Verification-grade memory for Claude Code. Every retrieved fact carries a staleness verdict (calendar age + filesystem path drift + git commit drift) so the model knows when a stored memory has rotted before relying on it; retrievals are logged with three attribution tiers (model-explicit with claim-level excerpts, Stop-hook substring match, or auto-fallback) so a reply months later can be traced back to the load-bearing stored claim wherever it was recorded. Memories are plain markdown + YAML in ~/.claude-memory/; retrieval is opt-in per turn; writes about you always stage for confirmation. A sibling episode tier (4 episode_* tools) gives /loop iterations and subagent handoffs a home for journal-shaped run-state the durability gate rejects. Bundles 25 MCP tools — 18 registered by default, the seven curation/power-user tools behind full_tool_surface (21 memory_* + 4 episode_*) — and a memory-discipline skill that lands the policy in the system prompt.