// the find
moorcheh-ai/memanto
Memory that AI Agents Love!
Memanto is a persistent memory layer for AI agents, exposing three operations — remember, recall, answer — backed by a proprietary semantic engine called Moorcheh that claims zero-latency indexing without vector DBs. It targets developers building with LangGraph, CrewAI, or Claude Code who are tired of agents losing context between sessions. The on-prem path runs via Docker; there's also a free cloud tier.
The three-primitive API (remember/recall/answer) is a genuinely cleaner abstraction than the blob-injection pattern most agent frameworks use. The 13 typed memory categories (fact, decision, goal, preference, etc.) give retrieval real filtering surface rather than treating all memories as undifferentiated text. The MCP integration means it drops into any MCP-compatible agent without custom glue code. Benchmark numbers on LongMemEval (89.8%) and LoCoMo (87.1%) are cited against real competitors with a public arXiv paper behind them, not just marketing claims.
The core retrieval engine (Moorcheh) is a black box — the 'information-theoretic search' claims are interesting but the engine itself isn't open source, so you're trusting a vendor for the most critical piece. The legacy/ directory is a visible scar: auth, idempotency, phase_d, safe_deletion, tracing modules all marked removed but still sitting in the tree, which suggests architectural churn that may not be done. The on-prem path requires Docker just to run the search backend, which is a real ops dependency for something marketed as zero-infrastructure. Session-based architecture means you have another stateful service to keep alive and monitor — the 'no backend to babysit' claim in the README conflicts with the memanto serve command and Docker requirement.