Everything since 0.0.3 — and the release that lines up with the redesigned README/docs.
Highlights
- Distillation dial (3 tiers, default no-LLM) — agent-driven distillation (the agent's own LLM, $0 to Midas) + an opt-in local distiller (
Memory(distiller=…),OllamaDistiller). Default off; measured honestly (naive distillation does not lift answers — see BENCHMARKS). - BEAM benchmark, all 4 tiers (100K→10M tokens) — recall@k 0.56→0.32 vs a 0.00 baseline; first judged answer-rate (BEAM-100K 0.40, LongMemEval ties LLM-ingest SOTA 0.84 — at $0 ingest).
- Bitemporal recall (
recall(as_of=…)), pinned standing directives, background auto-maintain, multilingual embedder option, opt-in ANN for huge stores. - TypeScript port (
npx -y midas-memory-mcp) — same tools/schema, shares one DB file live with the Python server. - Fix: belief revision now persists on SQLite-backed stores.
Install: pip install "midas-memory[local]" · uv tool install "midas-memory[mcp,local]" · npx -y midas-memory-mcp
Full changelog: CHANGELOG.md · benchmarks incl. the measured negatives: BENCHMARKS.md