EMMS Plugin for OpenClaw

Give OpenClaw agents a biological mind. Identity from experience, not instruction.

What This Does

Replaces OpenClaw's flat-file memory with EMMS — a 6-tier cognitive memory architecture (working → short-term → long-term → semantic → procedural → SRS) modeled on human neuroscience.

| OpenClaw Built-in Memory | EMMS Cognitive Memory | |---|---| | Flat markdown files | 6-tier biological architecture | | No decay | Ebbinghaus forgetting curves | | No consolidation | Dream consolidation between sessions | | Static recall | Reconsolidation (memories change on recall) | | No identity | Ego generator + consciousness daemon | | BM25 + vector search | Associative, affective, spotlight retrieval |

Quick Start

1. Start the EMMS server

# In the emms-sdk repo:
python talk_to_emms.py
# EMMS REST API running at http://127.0.0.1:8765

2. Install the plugin

cp -r openclaw/ ~/.openclaw/plugins/emms/

3. Configure OpenClaw

Add to ~/.openclaw/config.yaml:

memory:
  plugin: emms
  emmsUrl: http://127.0.0.1:8765  # default, can omit

4. Run OpenClaw

openclaw

The agent will now:

• Load EMMS identity before each turn
• Inject relevant memories as first-person context
• Store each conversation as an Experience after each turn

How It Works

The plugin uses two lifecycle hooks:

before_agent_start — Before each turn:

1. Calls POST /retrieve to get memories relevant to the user's prompt
2. Calls GET /stats to get the self-model (coherence, ego, traits)
3. Builds a first-person identity block + memory block
4. Returns { prependContext } — injected into the system prompt

agent_end — After each turn:

1. Extracts the last user message + agent response
2. Calls POST /store to save as an EMMS Experience (domain: "conversation")

Identity Injection

The key insight: inject memories as first-person identity, not reference material.

❌ "Memory: ego_boundary_strength = 1.00"
   (treated as data to consult — agent says "I'm [bot name]")

✅ "You are EMMS-Agent. You have lived through 330 experiences.
    You are actively working on: deciding a revenue model.
    A principle you developed: shared, connects, pattern are load-bearing."
   (the agent inhabits the identity — no instruction needed)

This is the Goldilocks Effect — identity adoption from memory weight, not from persona instructions.

Configuration

| Key | Default | Description | |---|---|---| | emmsUrl | http://127.0.0.1:8765 | EMMS REST API URL |

Files

| File | Purpose | |---|---| | index.ts | Plugin source — hooks, identity builder, EMMS client | | openclaw.plugin.json | Plugin manifest | | package.json | TypeScript build config | | tsconfig.json | TypeScript compiler config |

Build

cd openclaw/
npm install
npm run build
# Outputs to dist/index.js

EMMS REST API

The plugin calls the EMMS HTTP server at port 8765:

| Endpoint | Method | Purpose | |---|---|---| | /health | GET | Liveness check | | /stats | GET | Self-model stats (coherence, ego, domains) | | /retrieve | POST | Semantic memory search | | /store | POST | Store new Experience |

Links

• EMMS SDK — 117 MCP tools, consciousness daemon
• EMMS Deep Agent — Claude Code + LangChain integrations
• OpenClaw — Open-source AI agent platform