Memory OS launches six-layer local persistent Memory Stack for Hermes Agent

Memory OS, published under an MIT license and announced by developer ClaudioDrews on May 31 — June 1, 2026, delivers a six-layer persistent memory stack that runs locally alongside the Hermes Agent rather than as a toggleable plugin. The stack is intended to give Hermes — based workflows gated, on-device retrieval and long-term memory without relying on cloud memory services — a distinction that matters for teams with data-residency or subscription concerns.

The project is distributed as a standalone stack that preserves Hermes workspace files and the agent’s session database while adding additional persistence layers. The full system runs locally using Docker, Qdrant, Redis and Python 3.11+, and it interoperates with any LLM provider Hermes supports, including OpenRouter, OpenAI, Anthropic and Ollama. Memory OS is community — built and operates beside Hermes rather than as an official external provider. Memory OS defines six layered components: Layer 1 Workspace injects MEMORY.md, USER.md and CREATIVE.md into the system prompt; Layer 2 Sessions uses state.db (SQLite with FTS5 full-text search) for in-session state; Layer 3 Structured Facts stores memory_store.

and Layer 6 is an auto-curated LLM wiki fed continuously via wiki-continuous-ingest. Retrieval is organized around a pre_llm_call “surgical recall” that queries Fabric, Qdrant, Sessions and Facts in parallel. Each source is gated by a relevance threshold and per-session deduplication prevents duplicated context; a social — closer filter removes trivial messages (for example, plain “thanks”). The system also auto-extracts new learnings on post_llm_call and on_session_end to update memory stores while aiming for token efficiency.

Layer 5 search uses a four-step fallback cascade: hybrid search first, then dense vectors, then lexical methods, then SQLite, ensuring recall continues if one method returns nothing. Housekeeping features include a weekly decay scanner that ages stale entries and semantic deduplication that merges near-identical memories when cosine similarity exceeds 0.92, reducing long-term bloat. The README frames the project as a “memory operating system” positioned against cloud memory providers such as mem0, Zep and Letta by emphasizing a local — first, no-subscription approach. Hermes already supports multiple external memory providers (the project notes eight, including mem0 and Honcho); Memory OS remains a separate, community — maintained stack intended for teams that prefer local control and continuous, layered memory.