// the find
ClaudioDrews/memory-os
A 7-layer memory operating system for Hermes Agent — persistent memory with Qdrant, structured facts, fabric recall, auto-curated wiki, and surgical context injection. Runs locally, any LLM provider.
Memory OS is a multi-layer memory system bolted onto Hermes Agent, using Qdrant, SQLite, Redis, and ARQ to give it persistent, searchable memory across sessions. It runs entirely locally and works with any LLM provider Hermes supports. Aimed at power users who are serious about Hermes specifically — this won't help you if you're building on anything else.
The 4-level search fallback (hybrid Qdrant → dense → lexical → SQLite) is a sensible degradation strategy that prevents silent failures when the vector store is unavailable. The trust-scoring feedback loop on structured facts is a genuinely useful idea — facts the agent acts on successfully gain weight, stale ones decay. The 'Ground Truth hierarchy' framing (Layer 7) correctly identifies a real problem: injected context gets ignored without explicit instruction to treat it as authoritative. The one-command installer with smoke tests is a concrete commitment to actually running on real hardware, not just in demos.
This is tightly coupled to Hermes Agent — the README says 'Hermes Agent + Docker' as a hard requirement, meaning the entire architecture is only useful if you've already bought into that one agent framework. The bundled Icarus fork is explicitly not upstream-compatible, so you're taking on diverged maintenance debt the moment you install this. The 7-layer architecture description reads like design documentation, but there's no explanation of what breaks when layers conflict — what happens when Qdrant says one thing and the fact store says another? At 1102 stars with v0.2.0 still in progress and 20+ fixes from a 'systematic community audit,' this is not production-stable infrastructure yet.