Introducing Memory OS: A Revolutionary 6-Layer Open-Source Memory Stack for the Hermes Agent
A new project called Memory OS has been released, enhancing the capabilities of the Hermes Agent developed by Nous Research. Released under an MIT license by developer Claudio Drews, Memory OS introduces a six-layer memory architecture designed to enhance user interactions by enabling memory retention across sessions. This innovative library includes features such as a vector database, structured facts, and an auto-curated knowledge wiki.
Understanding Memory OS
Memory OS functions as a comprehensive system that operates alongside the native memory features of Hermes. While Hermes offers components like workspace files and session databases, Memory OS adds four additional layers. The entire stack runs locally, utilizing technologies such as Docker, Qdrant, Redis, and Python 3.11+. It is compatible with any large language model (LLM) provider supported by Hermes, including OpenRouter, OpenAI, Anthropic, and Ollama. The architecture is framed as a "memory operating system," highlighting its extensive functionality.
The Six Layers of Memory OS
- Layer 1: Workspace - This layer includes MEMORY.md, USER.md, and CREATIVE.md files. They all come together in the Memory OS system prompt during each interaction.
- Layer 2: Sessions - Using state.db, a SQLite database with full-text search, this layer keeps track of conversation history within the Memory OS framework.
- Layer 3: Structured Facts - Here, we store reliable facts in memory_store.db. It uses SQLite, HRR, FTS5, and trust scoring, with a feedback loop that constantly updates trust scores and resolves entities in Memory OS.
- Layer 4: Fabric - This layer is a tweaked version of the Icarus Plugin, improving session extraction through LLM-powered tools, allowing for cross-session recall in the Memory OS environment.
- Layer 5: Vector Database - Built on Qdrant, this layer uses 4096-dimensional cosine vectors and BM25 sparse search to rank keywords effectively in Memory OS.
- Layer 6: LLM Wiki - Think of this as an auto-curated repository of concepts, entities, and comparisons. It’s continuously updated and fed back into Qdrant through a process called wiki-continuous-ingest.
Managing Fallbacks and Cleanup in Memory OS
Memory OS has a solid four-level fallback system built into its memory retrieval process. It starts with hybrid search, then moves on to dense vectors, lexical searches, and finally SQLite if the earlier methods don’t work. This setup ensures memory retrieval stays effective, even if there’s a hiccup with the vector database. Plus, Memory OS runs a weekly decay scanner to eliminate outdated entries and uses smart semantic deduplication techniques to merge nearly identical memories when cosine similarity exceeds 0.92. This way, we can avoid memory bloating over time and keep everything running smoothly.
Strengths and Limitations of Memory OS
Strengths of Memory OS:
- It has a clear layered architecture that separates files, sessions, facts, vectors, and wiki elements, which greatly aids in memory management.
- Memory OS runs completely locally, so you don’t have to rely on cloud subscriptions—this keeps your data private and under your control.
- Its provider-agnostic design fits perfectly with Hermes Agent’s flexibility, making it easy to integrate with different systems.
- The retrieval processes are token-efficient, backed by gated sources and per-session deduplication for improved performance.
Limitations of Memory OS:
- It's still in its early stages, and the limited number of commits is affecting its development.
- The use of a forked Icarus Plugin means it’s not compatible with the upstream version, which may limit some functionalities.
- Setting it up can be a bit tricky—you’ll need Docker, Qdrant, Redis, and an ARQ Worker for it to function properly.
- There aren’t any published benchmarks on recall quality, latency, or token efficiency yet, which could make users somewhat hesitant.
Key Takeaways about Memory OS
- Memory OS is a community-driven, MIT-licensed memory stack that adds six extra memory layers to Hermes Agent.
- It brings together various components like workspace files, FTS5 session search, trust-scored facts, a forked Icarus fabric, Qdrant vectors, and an auto-curated LLM wiki.
- The retrieval process works on pre_llm_call, pulling gated recall from multiple sources, while new learnings are captured during post_llm_call and on_session_end.
- Its memory infrastructure is fully local and adaptable to different providers, although LLM calls are directed to the provider you choose.