ascottbell/maasv

Implements a full memory lifecycle with six stages—extract, store, consolidate, retrieve, decay, and forget—operating on a single SQLite database with local embeddings (Qwen3 via Ollama). Retrieval fuses three signals (dense vectors, BM25 FTS5, and knowledge graph connectivity) via Reciprocal Rank Fusion, optionally reranked by a learned 81-parameter neural network trained on actual retrieval patterns with Inverse Propensity Scoring. Integrates with any LLM provider (Anthropic, OpenAI) for automatic entity extraction, accepts data from HTTP or MCP sources, and tracks provenance across all memories—enabling shared cognition across multiple AI agents and tools.

Used by 1 other package. Available on PyPI.