Post-Cortex – Persistent memory for AI assistants with local semantic search (github.com)

Post-Cortex is a production-grade, Rust-built MCP (Model Context Protocol) server that turns ephemeral assistant sessions into persistent, private memory: conversations are auto-indexed, entity-extracted, and organized into a dynamic knowledge graph that can be queried by natural language. It runs entirely on-device (no external APIs) using local transformer models via the Candle framework, enabling semantic search, cross-session recall, and automatic context tracking for decisions, code changes, bugs and more — all exposed through 20+ MCP tools (create_session, load_session, semantic_search_session, get_structured_summary, etc.) and integrations with Claude Desktop and Zed. Technically notable: a lock-free concurrent architecture in Rust provides zero-deadlock guarantees under load; data is tiered (hot/warm/cold) with RocksDB persistence; embeddings (MiniLM default, StaticSimilarityMRL for M-series, TinyBERT, BGESmall) are stored in an HNSW index for fast nearest-neighbor search; and results are ranked by a combined_score = 0.7×similarity + 0.3×importance with sensible filtering thresholds. The system has been stress-tested with hundreds of conversations and thousands of entities, offering production-scale local semantic search (50–200ms) and ergonomics for long-lived agent memory — a practical privacy-preserving solution for assistants that need durable, context-aware recall.