| Summary | Self-improving learning engine for Claude Code. Extracts signals from every session (corrections, successes, confusion), runs periodic reflections, and evolves behavioral frameworks through evidence tiers (hypothesis → observed → validated). Frameworks that keep working get promoted, bad ones get retired. 9 MCP tools, automatic hooks, phase-adaptive learning. Single dependency, local-only. npx claude-soul init --starter | Local-first memory for long-horizon AI agents with no LLM at ingest — $0 per message, zero data egress, verbatim source-traceable recall. Typed belief revision, selective forgetting, and a provenance guard that blocks memory-justified destructive actions unless user-confirmed. One midas init wires Claude Code, Cursor, Codex, and 6 more clients to one shared SQLite memory; local web inspector and hash-chained audit log. Benchmarked on LongMemEval/LoCoMo/BEAM with one-command repros — including published negatives. uvx midas-memory-mcp or npx -y midas-memory-mcp |
|---|