d33pmemory OpenClaw Plugin

Automatic long-term memory for AI agents. No prompt engineering. No manual memory management.

This plugin connects OpenClaw agents to d33pmemory — a hosted memory-as-a-service platform. Every conversation is automatically analyzed, facts are extracted, and relevant memories are injected into the agent's context before every response.

Quick Start

1. Install

openclaw plugins install @d33pmemory/openclaw-plugin

Or for development:

git clone https://github.com/eduarddriessen1/d33pmemory-openclaw
openclaw plugins install -l ./d33pmemory-openclaw

2. Configure

Add to your openclaw.json:

{
  "plugins": {
    "entries": {
      "d33pmemory": {
        "enabled": true,
        "config": {
          "apiKey": "dm_your_api_key_here",
          "autoIngest": true,
          "autoRecall": true
        }
      }
    }
  }
}

Restart the gateway:

openclaw gateway restart

That's it. The agent will now automatically remember everything about the user.

What Gets Remembered

d33pmemory extracts structured memories from conversations:

| Type | Examples | |------|----------| | Facts | Name, age, occupation, location | | Preferences | Likes, dislikes, communication style | | Relationships | Family, friends, colleagues | | Events | Things the user has done or plans to do | | Patterns | Behavioral habits, recurring themes |

Memories are stored with confidence scores and source attribution (stated vs inferred).

Configuration Options

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | required | Your d33pmemory API key (dm_xxx) | | apiUrl | string | https://api.d33pmemory.com | API base URL | | agentId | string | derived from session | See Multi-Agent Safety | | autoIngest | boolean | true | Auto-ingest every conversation turn | | autoRecall | boolean | true | Auto-recall memories on session bootstrap | | recallQuery | string | general context query | Custom recall query for bootstrap | | recallMaxResults | number | 10 | Max memories injected on auto-recall | | recallMinConfidence | number | 0.3 | Minimum confidence threshold | | source | string | "openclaw" | Source label for ingested interactions |

Manual Tools

When the plugin is loaded, these tools are available to the agent automatically:

d33pmemory_recall

Search long-term memory for facts about the user.

Search: "what are this user's dietary restrictions"
Search: "what did the user mention about their weekend plans"
Search: "user's work projects and tech stack"

d33pmemory_ingest

Manually store something important (beyond auto-ingest).

Store: "User prefers to be called Eduard, not Ed"
Store: "User is building a SaaS called d33pmemory"

These tools are called automatically by the agent when it needs to recall or store something — no human trigger needed.

Multi-Agent Safety

OpenClaw supports multiple agents and workspaces. The plugin isolates each agent's memories correctly.

How isolation works

Every OpenClaw session has a session key:

agent:<workspace-name>:<channel>:<account>:<conversation>

Examples:

• agent:dm-agent:telegram:dm-agent-bot:direct:176654117
• agent:alice:telegram:alice-bot:direct:123456
• agent:dev-bot:slack:dev-bot:channel:C05ABC123

The plugin uses this to scope memory:

| Config agentId | agent_id in API | Effect | |---|---|---| | "my-team" (set) | "my-team" | All workspaces share memories | | "" or absent | Derived from session key | Each workspace is isolated |

Default behavior (no agentId set):

• dm-agent workspace → agent_id = "dm-agent"
• alice workspace → agent_id = "alice"
• Each workspace retrieves only its own memories

With shared agentId:

• Set "agentId": "team-shared" in config
• All workspaces share the same memory namespace
• Useful for team agents that should share context

Session-level deduplication

Each ingest also sends a custom_id derived from the full session key. This prevents one session's memories from overwriting another's, even within the same agent_id namespace.

How It Works

See ARCHITECTURE.md for the full technical explanation.

Auto-ingest flow

User message
    → Agent responds
        → "agent_end" lifecycle hook fires
            → Plugin extracts user turn + agent turn
            → resolveAgentId() → agent_id
            → POST /v1/ingest
                → d33pmemory extracts structured memories
                → Stored with session-scoped custom_id

Auto-recall flow

Session starts (/new)
    → "agent:bootstrap" hook fires
        → POST /v1/recall (scoped to workspace agent_id)
            → Relevant memories returned
            → Injected as D33PMEMORY_CONTEXT.md
    → Agent bootstrap context includes memories
        → Agent sees user context before first response

Repository Structure

d33pmemory-openclaw/
├── index.ts              # Plugin entry — hooks, tools, client
├── openclaw.plugin.json  # Plugin manifest + config schema
├── package.json
├── README.md             # This file
└── docs/
    └── ARCHITECTURE.md   # Technical deep-dive

Prerequisites

1. OpenClaw — Install guide
2. d33pmemory account — Sign up
3. d33pmemory API key — Create in your dashboard

Troubleshooting

Plugin not loading

openclaw plugins list
# Should show: d33pmemory ✓

Check the gateway log:

tail -f ~/.openclaw/gateway.log | grep d33pmemory

Hooks not firing

1. Make sure autoIngest / autoRecall are not explicitly set to false in config
2. Restart the gateway after config changes: openclaw gateway restart
3. Check openclaw status for plugin load confirmation

Memories not being recalled

1. Verify the apiKey is correct
2. Check that the agent has an entry in your d33pmemory dashboard
3. Try the d33pmemory_recall tool manually with a query like "user preferences and facts"

Wrong agent namespace

If memories are mixing between workspaces, check that agentId is not set in config if you want per-workspace isolation.

License

MIT