OpenClaw

OpenClaw is a personal AI assistant that runs locally and supports 11+ messaging platforms (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, Matrix, Zalo, etc.). Unlike all other AutoMem integrations, OpenClaw bypasses the MCP protocol entirely and calls the AutoMem HTTP API directly via curl commands through a native skill file.

---

Why Direct HTTP Instead of MCP

Aspect	OpenClaw	Other Platforms
Protocol	Direct HTTP API via curl	MCP over stdio
Integration type	Native skill (.md file)	MCP server process
Dependencies	bash + curl (built-in)	Node.js + npx
API access	Direct REST endpoints	MCP tool abstraction
Authentication	HTTP headers	Environment variables passed to MCP server

This approach eliminates the MCP abstraction layer, making the integration simpler and more reliable for bot-based systems where the AI executes shell commands directly.

---

Multi-Platform Memory Flow

graph TB
    subgraph "Messaging Platforms"
        WA["WhatsApp"]
        TG["Telegram"]
        SL["Slack"]
        DC["Discord"]
        SG["Signal"]
        IM["iMessage"]
        OT["Teams, Matrix, Zalo, etc."]
    end

    subgraph "OpenClaw Gateway"
        OC["OpenClaw Process"]
        SK["~/.openclaw/skills/automem/SKILL.md"]
    end

    subgraph "Memory Storage"
        LOCAL["Local: memory/YYYY-MM-DD.md"]
        CURATED["Local: MEMORY.md"]
        AUTOMEM["AutoMem: FalkorDB + Qdrant"]
    end

    WA --> OC
    TG --> OC
    SL --> OC
    DC --> OC
    SG --> OC
    IM --> OC
    OT --> OC

    OC --> SK
    SK -->|"curl $AUTOMEM_ENDPOINT"| AUTOMEM

    OC --> LOCAL
    OC --> CURATED

The OpenClaw gateway receives messages from all platforms, loads skills (including automem/SKILL.md), and the bot executes curl commands to interact with the AutoMem service. This provides persistent semantic memory across all messaging platforms.

---

Installation

CLI Installer

# Basic installation (auto-detects workspace)
npx @verygoodplugins/mcp-automem openclaw

# With explicit workspace and endpoint
npx @verygoodplugins/mcp-automem openclaw \
  --workspace ~/clawd \
  --endpoint https://your-automem.up.railway.app \
  --api-key your-token-here

# Preview without writing
npx @verygoodplugins/mcp-automem openclaw --dry-run

CLI options:

Option	Description	Default
--workspace <path>	OpenClaw workspace directory	Auto-detected
--endpoint <url>	AutoMem service endpoint	http://127.0.0.1:8001
--api-key <key>	AutoMem API key	None (optional)
--name <name>	Project name for memory tags	Auto-detected
--dry-run	Preview changes without modifying files	Off
--quiet	Suppress output	Off

What the installer does:

1. Locates OpenClaw workspace directory
2. Installs SKILL.md to ~/.openclaw/skills/automem/SKILL.md
3. Updates ~/.openclaw/openclaw.json with AutoMem environment variables
4. Creates <workspace>/memory/ directory for daily log files
5. Removes legacy AGENTS.md blocks from previous installations

Workspace Detection

The installer searches for the workspace in this order:

1. --workspace flag (explicit)
2. OPENCLAW_WORKSPACE or CLAWDBOT_WORKSPACE environment variable
3. ~/.openclaw/openclaw.json config file (reads agents.defaults.workspace)
4. Default paths: ~/.openclaw/workspace, ~/clawd, ~/.clawdbot/workspace

File Structure After Installation

~/.openclaw/
├── openclaw.json          # Updated with AutoMem config
└── skills/
    └── automem/
        └── SKILL.md       # Skill file with API reference

<workspace>/
└── memory/
    ├── .gitkeep
    └── YYYY-MM-DD.md      # Daily memory files (created by OpenClaw)

---

Configuration

The installer updates ~/.openclaw/openclaw.json under skills.entries.automem:

{
  "skills": {
    "entries": {
      "automem": {
        "env": {
          "AUTOMEM_ENDPOINT": "http://127.0.0.1:8001",
          "AUTOMEM_API_KEY": "your-token-here"
        }
      }
    }
  }
}

The installer uses deep merging — it preserves all existing configuration and only updates the automem entry.

Note

The installer handles JSON5-style comments in OpenClaw config files (single-line //, block /* */, trailing commas) using a string-aware comment stripper that preserves URLs containing //.

---

SKILL.md Architecture

The skill file at ~/.openclaw/skills/automem/SKILL.md uses YAML front matter to declare required environment variables:

---
env_required:
  - AUTOMEM_ENDPOINT
env_optional:
  - AUTOMEM_API_KEY
---

The rest of the file contains the complete HTTP API reference that the bot uses to construct curl commands.

Memory Operations

Store a memory:

curl -s -X POST "$AUTOMEM_ENDPOINT/memory" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AUTOMEM_API_KEY" \
  -d '{
    "content": "Brief title. Context and details. Impact/outcome.",
    "type": "Decision",
    "importance": 0.9,
    "tags": ["project-name", "component", "YYYY-MM"]
  }'

Recall memories:

curl -s "$AUTOMEM_ENDPOINT/recall?query=your+search+query&limit=5&tags=project-name" \
  -H "Authorization: Bearer $AUTOMEM_API_KEY"

Recall parameters: query, limit, tags, tag_mode, time_query, expand_entities.

Associate memories:

curl -s -X POST "$AUTOMEM_ENDPOINT/associate" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AUTOMEM_API_KEY" \
  -d '{
    "memory1_id": "id-one",
    "memory2_id": "id-two",
    "type": "LEADS_TO",
    "strength": 0.8
  }'

11 relationship types: RELATES_TO, LEADS_TO, EVOLVED_INTO, DERIVED_FROM, INVALIDATED_BY, CONTRADICTS, REINFORCES, PREFERS_OVER, PART_OF, EXEMPLIFIES, OCCURRED_BEFORE.

---

Behavioral Rules

Session Start — Recall First

The skill instructs the bot to recall at session start for:

• Questions about past decisions, preferences, or history
• Debugging or troubleshooting (search for similar past issues)
• Project planning or architecture discussions

Skip recall for:

• Simple greetings or small talk
• Questions answerable from general knowledge
• Direct file operations or commands

Storage Importance Levels

Category	Importance	Examples
Decisions	0.9	”Chose Railway over Fly.io for deployment. Reason: persistent volumes.”
User corrections	0.8	”Human prefers dark mode themes. Corrected my light mode suggestion.”
Bug fixes	0.8	”WhatsApp webhook failing. Root cause: expired token. Solution: auto-refresh.”
Preferences	0.7	”Human likes terse responses, no fluff.”
Patterns	0.7	”Use early returns for validation in all API routes.”
Context	0.5	”Set up new Telegram channel for family group.”

Three Memory Layers

OpenClaw uses three complementary memory layers:

1. Daily files (memory/YYYY-MM-DD.md) — raw conversation logs, created automatically
2. MEMORY.md — curated notes in the workspace
3. AutoMem (FalkorDB + Qdrant) — semantic search and graph relationships across all sessions and platforms

AutoMem provides the cross-session, cross-platform memory that the file-based layers cannot.

---

Error Handling

The skill instructs the bot on graceful degradation:

• If AutoMem is unavailable (curl fails): Continue normally — memory enhances but never blocks
• Do not announce failures to the human
• Fall back to file-based memory (memory/ directory and MEMORY.md)
• Only check /health endpoint to diagnose persistent failures

Common curl Error Patterns

Error	Cause	Solution
401 Unauthorized	Missing Authorization header	Add -H "Authorization: Bearer $AUTOMEM_API_KEY"
Connection refused	AutoMem service not running	Start service or fall back to file-based memory
Empty results	Query too specific or no matches	Broaden query terms, remove tag filters, increase limit

---

Troubleshooting

Agent not using memory

1. Verify skill is loaded: openclaw skills check automem
2. Restart OpenClaw gateway after installation
3. Check AutoMem service is running: curl http://127.0.0.1:8001/health

Bot mentions “Memory Tools Disabled”

This refers to OpenClaw’s built-in memory-lancedb plugin, not AutoMem. The skill explicitly instructs the bot to ignore this message — AutoMem handles embeddings server-side with no client API keys required.

curl calls failing

1. Verify endpoint: curl $AUTOMEM_ENDPOINT/health
2. Check API key if using an authenticated instance
3. Check firewall/VPN for Railway endpoints
4. Examine OpenClaw logs for curl exit codes

---

Comparison with MCP Integrations

Feature	OpenClaw (Direct HTTP)	MCP Platforms
Protocol	REST API via curl	MCP over stdio
Setup complexity	Single CLI command	CLI + config file editing
Runtime dependencies	bash + curl	Node.js runtime
Error recovery	curl exit codes	MCP error protocol
Latency	Direct HTTP call	MCP serialization overhead
Platform specificity	Native skill system	Universal MCP
Auth method	HTTP Bearer token	Environment variables

The direct HTTP approach is simpler for bot-based systems but lacks the standardization benefits of MCP for cross-platform integrations.