Memory OS

Memory OS is an open-source, MIT-licensed memory operating system built by Claudio Drews for Hermes Agent. It runs entirely on your local machine and works with any LLM provider — OpenRouter, OpenAI, Anthropic, Ollama, or local models — with no cloud memory subscription or vendor lock-in. The project was created to solve the persistent problem of agent amnesia: losing context, decisions, and structured knowledge between sessions.

What It Is

Memory OS is a complete local memory infrastructure layer for Hermes Agent, not a simple plugin. It stacks six distinct memory layers — from flat markdown files to a Qdrant vector database — that work in concert to give an agent genuine long-term recall. The system automatically injects the right context at the right moment using relevance-gated retrieval, per-session deduplication, and a social-closer filter that skips trivial messages. The result is what the README describes as "surgical token efficiency": the LLM receives exactly what it needs, nothing more.

The Six-Layer Architecture

Each layer handles a different scope and type of memory:

• Layer 1 – Workspace: Markdown files (MEMORY.md, USER.md, CREATIVE.md) injected into the system prompt every turn
• Layer 2 – Sessions: SQLite with FTS5 full-text search across the entire conversation history
• Layer 3 – Structured Facts: A dedicated SQLite store with entity resolution, trust scoring, and an automatic feedback loop that trains trust scores over time
• Layer 4 – Fabric (Cross-Session): A heavily forked version of the Icarus plugin with LLM-powered session extraction and multi-source injection across 16 tools
• Layer 5 – Vector Database: Qdrant running locally with 4096-dimensional cosine similarity plus BM25 sparse search, a 4-level fallback cascade, weekly decay scanning, and semantic deduplication (cosine > 0.92 triggers a merge)
• Layer 6 – LLM Wiki: A self-curating knowledge vault organized into concepts/, entities/, and comparisons/ directories, continuously ingested into Qdrant

The flow is: pre_llm_call triggers surgical recall from all four retrieval sources; post_llm_call and on_session_end trigger automatic learning extraction and capture.

Infrastructure and Requirements

Memory OS requires Hermes Agent, Docker (for Qdrant, Redis, and an ARQ Worker), and Python 3.11+. The entire memory stack runs locally — no data leaves the machine. The bundled Icarus fork (icarus/) is not upstream-compatible with the original esaradev/icarus-plugin; key additions include LLM-powered session extraction, multi-source injection, CREATIVE.md isolation, backtick sanitization, a system injection filter, and social closer detection. A companion project, Vault Curator v3 (ClaudioDrews/vault-curator), handles frontmatter enrichment, semantic linking, and MOC index generation for the wiki layer.

How It Differs from Cloud Memory Solutions

The README explicitly positions Memory OS against cloud-first alternatives like mem0, Zep, and Letta, noting that those solutions require cloud memory subscriptions and do not offer full local infrastructure. Memory OS's stated differentiators include: fully local memory infrastructure, no memory subscription, provider-agnostic LLM support, Hermes-native integration, structured facts with trust scores, a self-curating wiki, and intelligent decay with archival — none of which the README attributes to the compared alternatives.

Current Status

The repository was created on May 31, 2026, and last pushed on June 1, 2026, making it a very recent release. It is written primarily in Python, licensed under MIT, and hosted at github.com/ClaudioDrews/memory-os. At the time of indexing the project had 38 stars and 1 fork with no open issues.