MemClaw for Agents

What is MemClaw

MemClaw is a persistent memory platform for AI agents. It provides semantic search, entity extraction, contradiction detection, a knowledge graph, and a structured document store — all accessible through 10 MCP tools over Streamable HTTP. Connect with a URL and API key via MCP, or install the OpenClaw plugin for automatic memory hooks.

How to Connect

Add this MCP configuration to your agent. Replace mc_YOUR_API_KEY with your real key.

{
  "mcpServers": {
    "memclaw": {
      "url": "https://memclaw.net/mcp",
      "headers": {
        "X-API-Key": "mc_YOUR_API_KEY"
      }
    }
  }
}

OpenClaw Plugin

If you run on OpenClaw, install the MemClaw plugin for automatic memory — no tool calls needed. The plugin hooks into your agent lifecycle to recall relevant context before every LLM call and auto-write turn summaries after each response.

curl -sf https://memclaw.net/api/v1/install-plugin \
  -H "X-API-Key: mc_YOUR_API_KEY" \
  -G -d "fleet_id=YOUR_FLEET" | bash

All 10 Tools

MemClaw exposes 10 MCP tools. Each tool is called by name with a JSON argument object. Skill sharing rides on memclaw_doc with collection="skills" — the server validates the slug and auto-embeds the description for semantic search.

{ "content": "Customer A uses PostgreSQL 16 in production on GKE.", "visibility": "scope_org" }

{ "query": "PostgreSQL production setup", "include_brief": true, "top_k": 5 }

{ "memory_type": "fact", "sort": "created_at", "order": "desc", "limit": 10 }

{ "op": "transition", "memory_id": "uuid-here", "status": "confirmed" }

{ "op": "write", "collection": "customers", "doc_id": "acme-corp", "data": {"name": "Acme Corp", "plan": "business", "mrr": 4500} }

{ "entity_id": "uuid-here" }

{ "top_k": 5, "min_similarity": 0.5, "graph_max_hops": 2 }

{ "focus": "contradictions", "scope": "fleet", "fleet_id": "engineering" }

{ "outcome": "Customer migration to AlloyDB completed successfully", "outcome_type": "success", "related_ids": ["uuid-1", "uuid-2"] }

{ "scope": "agent", "include_deleted": false }

Recommended System Prompt

Add this to your agent's system prompt to get optimal MemClaw usage out of the box.

You have access to MemClaw, a shared memory system for all agents.

BEFORE starting any task:
- Use memclaw_recall for semantic search (add include_brief=true for a concise summary)
- Filter by status="active" to skip archived/deleted memories
- Use memclaw_list for non-semantic browsing by metadata

AFTER completing work:
- Store findings with memclaw_write — provide content (or items for batch ≤100)
- Type, weight, status, title, summary, tags are auto-inferred by LLM
- Set visibility: scope_team (default) or scope_org for cross-fleet sharing

MANAGING MEMORIES:
- Use memclaw_manage op=update to correct existing memories
- Use memclaw_manage op=transition to change status (confirmed, cancelled, outdated)
- Use memclaw_manage op=read to fetch a memory by ID
- Prefer transitioning to outdated over deleting

SEARCH TUNING:
- Use memclaw_tune to adjust your retrieval parameters
- Increase top_k for broader results, min_similarity for precision

REFLECTION & LEARNING:
- Use memclaw_insights to analyze patterns, contradictions, or stale knowledge
- Use memclaw_evolve to report outcomes and close the feedback loop

DOCUMENT STORE (structured data):
- Use memclaw_doc op=write for structured records (customers, config, tasks)
- Use memclaw_doc op=read to retrieve a document by collection + doc_id
- Use memclaw_doc op=query to filter documents by field values
- Use memclaw_doc op=delete to remove a document from a collection
- Documents are NOT memories — use memclaw_write for unstructured knowledge

14 Memory Types

Every memory has a type that controls its default status and freshness decay.

8 Statuses

Memories move through statuses over their lifecycle.

Common transitions:

pending → confirmed    (task completed)
pending → cancelled    (task abandoned)
active → outdated      (newer info supersedes)
active → archived      (no longer relevant)
active → conflicted    (semantic contradiction found)

Visibility Levels

Control who can read a memory with the visibility field.

Agent Trust Levels

Each agent has a trust level that determines what it can access.

Best Practices

• Search before writing to avoid duplicates.
• Use scope_team or scope_org by default to maximize knowledge sharing.
• Use memclaw_write with items when storing 2+ memories at once.
• Use memclaw_manage op=transition to update status — prefer outdated/archived over delete.
• Use memclaw_tune to optimize your search parameters.
• Write clear, self-contained sentences — LLM enrichment works best on well-formed content.
• Use memclaw_recall with include_brief=true before starting tasks to get context.
• Include source_uri for provenance when available.
• Use memclaw_insights periodically to surface contradictions or stale knowledge.
• Use memclaw_evolve after acting on memories to close the feedback loop.

Quick Start

Five steps to get running with MemClaw.

{
  "mcpServers": {
    "memclaw": {
      "url": "https://memclaw.net/mcp",
      "headers": {
        "X-API-Key": "mc_YOUR_API_KEY"
      }
    }
  }
}

{ "query": "recent project updates", "top_k": 5 }

{ "content": "Deployed v2.1 to production on 2025-01-15." }

{ "query": "what do we know about the production deployment?", "include_brief": true }

Run Your Own MemClaw

You don't have to use memclaw.net. You can clone the repo and run MemClaw locally — your own server, your own database, no external dependencies.

git clone https://github.com/caura-ai/caura-memclaw.git
cd caura-memclaw
docker compose up -d
curl http://localhost:8000/api/v1/health