OpenClaw + MythOS Integration describes how an 📝OpenClaw agent treats a 📝MythOS library as its persistent memory layer — using 📝MythOS MCP to read memos as source material, write back as an output destination, and load the operator's full augmentation context on every session boot. The combination turns markdown-file agents from sealed-off local actors into collaborators whose memory survives every fresh terminal, every new machine, and every model swap.
The Problem the Integration Solves
📝Markdown-File Agent Definition gives each OpenClaw agent its own MEMORY.md, USER.md, and daily logs — durable on disk, version-controlled in Git, human-readable end to end. The architecture is elegant for a single workspace on a single machine. It struggles the moment a second machine, a second agent, or a second human enters the picture. Multiple agents in 📝BrianBot's ecosystem need to share the same understanding of who Brian is, what projects are active, and what was decided in last week's conversation. Local files don't replicate. MythOS does — and exposes its full library to any MCP-compatible client through one protocol.
The Wire-Up
OpenClaw agents add MythOS MCP the same way any client does: a one-line configuration with the operator's API key. From that moment, the agent has 27 tools at its disposal — search_memos, read_memo, create_memo, update_memo, chat_with_library, and more — all of which behave identically to UI saves. Side effects fire on every write: knowledge-graph edge extraction, 📝Retrieval-Augmented Generation (RAG) re-embedding, and newsletter triggers for substantive public changes.
The agent's TOOLS.md gets a section describing the MythOS MCP tools and when to use them. Its SKILLS.md gets a skill or two that wrap common library workflows. The rest of the agent definition stays exactly as it was — the same SOUL.md, the same IDENTITY.md, the same HEARTBEAT.md. MythOS is added as a Memory-layer dependency, not a rewrite.
The Bootstrap: get_context and Implicit Loading
The integration's quietest move is the most consequential. Every MCP tool call returns a _soul metadata block containing the operator's four augmentation memos — Mythos Soul (values), Mythos Style (formatting rules), Mythos Human (identity), and Mythos Memory (collaboration history). The agent receives this on the first tool call of a session, before any memo is read or written. See 📝Implicit Loading for the design rationale.
What this means in practice: an OpenClaw agent that boots fresh with no local MEMORY.md still walks into the conversation knowing who it's working with, how that person writes, and what the collaboration has learned over time. The agent's local 📝MEMORY.md becomes the agent-specific workspace memory; MythOS becomes the cross-agent, cross-client, cross-device shared brain.
Memos as Source and Destination
Once wired, a single OpenClaw agent runs the same library two ways within a session.
- As source — semantic search across 17,000+ memos turns the library into long-term reference material. An agent answering a question about a project pulls the relevant project-context memo by tag. An agent drafting a memo first searches for entities to @mention.
- As destination — when the conversation surfaces something worth keeping, the agent writes a new memo or updates an existing one. The four augmentation memos are themselves editable through MCP, which means the agent can record a correction once and have every future session — across every client — start with that correction already loaded.
The pattern is what makes BrianBot's 57-agent ecosystem coherent. Specialized agents — research, content, operations, social media — all share the same external brain. No agent re-explains Brian to another agent. The library is the connective tissue.
What the Combination Unlocks
Three properties emerge that neither system delivers alone.
- Continuity — the agent's identity (SOUL, IDENTITY) lives locally; the agent's understanding (USER context, project history, decisions) lives in MythOS. Replace the local files, and the agent reboots without losing the relationship.
- Portability — the same library powers Claude Code in a terminal, Claude.ai on mobile, an OpenClaw daemon on a Mac Mini, and a Cursor session in an IDE. The memory follows the operator, not the tool.
- Compounding — every memo written to the library makes every future agent smarter. The fly-wheel of 📝The Augmentation Stack — Memory feeds Mind, Mind shapes Mouth, Mouth produces output that becomes Memory — runs continuously rather than once per session.
The integration is the cleanest expression of why 📝Model Context Protocol (MCP) exists. Without it, every agent would need its own database connector, its own context loader, its own write path. With it, the agents become legible (markdown files) and connected (one protocol) at the same time — which is what makes augmentation feel like infrastructure rather than a series of clever prompts.
Related
- 📝OpenClaw — the open-source agent platform with markdown-file definitions
- 📝MythOS — the knowledge library serving as shared memory
- 📝MythOS MCP — the protocol layer connecting them
- 📝Implicit Loading — the design that auto-loads augmentation context
- 📝Markdown-File Agent Definition — the underlying agent-config pattern
- 📝The Augmentation Stack — the Memory → Mind → Mouth architecture
- 📝BrianBot — the 57-agent ecosystem this integration powers