pmm vs. total-recall

If you have ever had AI hallucinate and make things up (and gaslight you, even): it's because summaries do not preserve nuance. Each /compact cycle summarises and compresses. Decisions lose their rationale. Lessons lose their context. By Monday, you're re-explaining choices you made on Thursday. That's context rot. PMM fixes that. Every session, Claude loads what it already knows, then adds to it. Markdown files, git-backed. No database, no setup. Decisions tracked with rationale: not just what but. git log as your memory audit trail. Memory that compounds, not rots.

Coding Agents forget everything when a session ends. total-recall fixes that: a shared memory layer that persists across sessions, tools, and devices. Memory is organized into three tiers (Hot / Warm / Cold) with a 4,000-token budget enforced automatically — so Claude gets relevant context injected at session start without carrying stale noise. Semantic search (all-MiniLM-L6-v2 embeddings, offline) retrieves the right memories on demand. A knowledge base lets you ingest READMEs, API docs, and architecture notes that are retrieved semantically when relevant. An eval framework lets you measure retrieval quality, run benchmarks, and compare config changes — so you know memory is actually helping. All state is local: SQLite + ONNX embeddings, no external services, no API keys required. Works across Claude Code, Copilot CLI, Cursor, Cline, and OpenCode — existing memories auto-import on first run.