openclaw-supermemory

OpenClaw plugin for full supermemory-driven memory management.

Your AI assistant remembers everything — every session, every conversation. No more context loss.

What it does

• Auto-Recall: Before every response, searches supermemory for relevant context and injects it into the prompt (via before_prompt_build hook)
• Auto-Capture: After every conversation, sends ALL messages (user + assistant) to supermemory — supermemory decides what to remember
• Compaction Hook: Automatically saves conversation content before session compaction
• Agent Tools: Registers memory_save, memory_recall, memory_forget as native agent tools (no manual mcporter calls needed)

Prerequisites

• OpenClaw (2026.3+ recommended)
• mcporter CLI installed and configured with supermemory
• A supermemory account and API key

mcporter setup

# Install mcporter
npm install -g mcporter

# Configure supermemory MCP server (~/.mcporter/mcporter.json)
cat > ~/.mcporter/mcporter.json << 'EOF'
{
  "mcpServers": {
    "supermemory": {
      "baseUrl": "https://mcp.supermemory.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${SUPERMEMORY_API_KEY}"
      }
    }
  }
}
EOF

# Set your API key
export SUPERMEMORY_API_KEY="your-key-here"

# Verify it works
mcporter call supermemory.recall query="test"

Install

Option A: Copy to extensions directory

git clone https://github.com/kandotrun/openclaw-supermemory.git
cp -r openclaw-supermemory ~/.openclaw/extensions/supermemory-auto

Option B: Load from custom path

Add to your openclaw.json:

{
  "plugins": {
    "load": {
      "paths": ["/path/to/openclaw-supermemory"]
    }
  }
}

Enable the plugin

{
  "plugins": {
    "entries": {
      "supermemory-auto": {
        "enabled": true,
        "config": {
          "autoRecall": true,
          "autoCapture": true
        }
      }
    },
    "slots": {
      "memory": "supermemory-auto"
    }
  }
}

Then restart the gateway:

openclaw gateway restart

Configuration

Option	Type	Default	Description
autoRecall	boolean	true	Search supermemory and inject context before each response
autoCapture	boolean	true	Send ALL conversation messages (user + assistant) to supermemory
recallTimeoutMs	number	10000	Timeout for supermemory recall calls (ms)
captureTimeoutMs	number	20000	Timeout for supermemory save calls (ms)
maxQueryLength	number	300	Max characters from user message used as search query
maxContextChars	number	15000	Max characters of recalled context injected into prompt
maxCaptureChars	number	5000	Max characters per capture payload sent to supermemory
containerTag	string	""	Optional supermemory project/container tag for scoping memories

How it works

User sends message
        │
        ▼
┌─────────────────────────┐
│ before_prompt_build      │  ← Recall phase
│                         │
│ 1. Extract query        │
│    (strip metadata)     │
│ 2. mcporter call        │
│    supermemory.recall   │
│ 3. Inject context       │
│    into prompt          │
└─────────────────────────┘
        │
        ▼
   Agent responds
   (with full context)
        │
        ▼
┌─────────────────────────┐
│ agent_end                │  ← Capture phase
│                         │
│ 1. Collect ALL messages │
│    (user + assistant)   │
│ 2. Label each:          │
│    [ユーザー] / [アシスタント] │
│ 3. Strip noise          │
│    (metadata, context)  │
│ 4. Send to supermemory  │
│    (it decides what     │
│     to remember)        │
└─────────────────────────┘

Session compaction triggered
        │
        ▼
┌─────────────────────────┐
│ before_compaction        │  ← Pre-compaction save
│                         │
│ Save full conversation  │
│ to supermemory before   │
│ messages are pruned     │
└─────────────────────────┘

Agent Tools

The plugin registers three tools that the agent can use directly:

Tool	Description
memory_save	Save information to supermemory (with optional containerTag)
memory_recall	Search supermemory for relevant memories
memory_forget	Remove specific information from supermemory

These tools use mcporter internally — the agent doesn't need to call mcporter manually.

Capture Format

Messages sent to supermemory are clearly labeled:

[ユーザー] 明日福岡に行くんだけど、充電スポット調べて

[アシスタント] 福岡エリアのTesla Supercharger一覧:
- 福岡天神SC (4基)
- LECT広島SC (8基)
...

Supermemory receives the full exchange and decides what's worth remembering. No client-side filtering — supermemory's AI handles the curation.

Skipped messages

The plugin skips:

• Heartbeat/system messages
• Messages shorter than 5 characters
• NO_REPLY / HEARTBEAT_OK responses
• Already-injected supermemory-context blocks
• Metadata blocks (JSON, conversation info)

Notes

• Latency: Each recall adds ~2-3s (mcporter process spawn + MCP call). This is the trade-off for always having context.
• No filtering by design: v2 sends everything to supermemory. Supermemory's AI decides what to remember. This avoids losing important context that pattern-based filters would miss.
• Memory slot: Setting plugins.slots.memory to supermemory-auto replaces the default memory-core plugin. If you want both, don't set the memory slot.
• containerTag: Use this to scope memories to a specific project (e.g., sm_project_default).

Changelog

v2.0.0 (2026-03-13)

• Full rewrite — supermemory-driven architecture
• Migrated from before_agent_start (legacy) to before_prompt_build
• Full conversation capture: user + assistant messages, labeled with [ユーザー] / [アシスタント]
• No client-side filtering — supermemory decides what to remember
• Added before_compaction hook for pre-compaction saves
• Registered agent tools: memory_save, memory_recall, memory_forget
• Added containerTag support for project-scoped memories
• Expanded maxBuffer to 1MB for large recall results
• Better query extraction (strips context blocks + metadata noise)

v1.0.0 (2026-03-12)

• Initial release
• Auto-recall via before_agent_start
• Optional auto-capture (user messages only, pattern-filtered)

License

MIT

Copyright (c) 2026 Kan Ninomiya & 白川 玲 (AI)