Shodh is different – it's a cognitive memory system: - Hebbian learning – memories recalled together strengthen their connections - Activation decay – unused memories fade (like biological memory) - Knowledge graphs – entities extracted and connected automatically - Three-tier storage – working → session → long-term (based on Cowan's working memory model)
Technical details: - Single ~15MB Rust binary, runs 100% offline - Sub-millisecond graph operations (entity lookup: 763ns, 3-hop traversal: 30µs) - MiniLM-L6 embeddings + TinyBERT NER, bundled - MCP server (works with Claude, Cursor), Python SDK, REST API - RocksDB for persistence
Edge deployment: Designed for local-first and air-gapped environments. Runs on Jetson, Raspberry Pi, industrial PCs, drones – anywhere you need on-device memory without cloud round-trips. Zero network dependency.
What it's not: Not a vector database. Not RAG. It's meant to give agents persistent memory that actually learns and consolidates over time.
GitHub: https://github.com/varun29ankuS/shodh-memory Website: www.shodh-rag.com/memory
Happy to answer questions about the architecture or cognitive models.