openclaw-backboard

OpenClaw memory plugin powered by Backboard.io — the world's highest-scoring AI memory system. Gives your AI agent persistent, semantic long-term memory that works across sessions, conversations, and time.

Why Backboard?

Most AI memory systems forget, hallucinate, or fail at multi-step reasoning over long histories. Backboard's stateful memory architecture is different — it achieved 90.1% overall accuracy on LoCoMo, the industry-standard benchmark for long-term conversational memory, setting the highest public score ever recorded.

LoCoMo Benchmark Results

| Method | Single-Hop | Multi-Hop | Open Domain | Temporal | Overall | |---|---|---|---|---|---| | Backboard | 89.36% | 75.00% | 91.20% | 91.90% | 90.10% | | Memobase (v0.0.37) | 70.92% | 46.88% | 77.17% | 85.05% | 75.78% | | Zep | 74.11% | 66.04% | 67.71% | 79.79% | 75.14% | | Memobase (v0.0.32) | 63.83% | 52.08% | 71.82% | 80.37% | 70.91% | | Mem0-Graph | 65.71% | 47.19% | 75.71% | 58.13% | 68.44% | | Mem0 | 67.13% | 51.15% | 72.93% | 55.51% | 66.88% | | LangMem | 62.23% | 47.92% | 71.12% | 23.43% | 58.10% | | OpenAI | 63.79% | 42.92% | 62.29% | 21.71% | 52.90% |

Evaluated with GPT-4.1 as the LLM judge under standardized conditions. Full logs, prompts, and verdicts published for reproducibility.
Replication scripts: Backboard-io/Backboard-Locomo-Benchmark

Backboard leads in every category — single-hop recall, multi-hop reasoning, open-domain knowledge, and temporal reasoning. These are the same dimensions measured by LongMemEval (ICLR 2025), which evaluates five core long-term memory abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. Backboard's architecture is purpose-built to excel at exactly these tasks.

Features

• 5 agent tools — memory_search, memory_store, memory_get, memory_list, memory_forget
• Auto-recall — relevant memories injected into agent context before each turn
• Auto-capture — key facts stored automatically after each agent turn
• Session + long-term scopes — session-scoped memory (ephemeral) and assistant-scoped memory (persistent)
• Semantic search — vector-based retrieval, no keyword matching
• CLI commands — openclaw backboard search and openclaw backboard stats
• Zero-config assistant — auto-creates a Backboard assistant if none is specified
• 2,200+ LLM compatible — works with any model Backboard supports

Prerequisites

• OpenClaw gateway installed
• A Backboard.io account and API key
• Node.js 22+

Installation

git clone <this-repo> ~/.openclaw/plugins/openclaw-backboard
cd ~/.openclaw/plugins/openclaw-backboard
npm install

Configuration

Add the plugin to your OpenClaw config (~/.openclaw/openclaw.json):

{
  "plugins": {
    "load": {
      "paths": ["~/.openclaw/plugins/openclaw-backboard"]
    },
    "entries": {
      "openclaw-backboard": {
        "enabled": true,
        "config": {
          "apiKey": "bb_your_api_key_here",
          "assistantId": "",
          "autoCapture": true,
          "autoRecall": true,
          "topK": 5
        }
      }
    },
    "slots": {
      "memory": "openclaw-backboard"
    }
  }
}

Config Options

| Option | Type | Default | Description | |---|---|---|---| | apiKey | string | (required) | Your Backboard API key | | assistantId | string | auto-created | Backboard Assistant ID. Leave empty to auto-create one named "OpenClaw Memory" | | autoCapture | boolean | true | Automatically store conversation context after each agent turn | | autoRecall | boolean | true | Automatically inject relevant memories before each agent turn | | topK | number | 5 | Maximum memories returned per search |

Usage

Start the Gateway

./start.sh

This clears local cache, installs dependencies, and starts the OpenClaw gateway with the plugin loaded.

Agent Tools

The plugin registers five tools that your AI agent can call directly:

memory_store

Save information to long-term memory.

Agent, please remember that my preferred language is Python and I deploy to AWS.

The agent calls memory_store with the text. Supports optional metadata and a longTerm flag (true for assistant-scoped, false for session-scoped).

memory_search

Semantic search across stored memories.

What do you remember about my deployment preferences?

Returns ranked results with relevance scores. Supports scope parameter: "session", "long-term", or "all".

memory_list

List all stored memories. Useful for auditing what the agent knows.

memory_get

Retrieve a specific memory by ID.

memory_forget

Delete a memory by ID or by search query. Single high-confidence matches are deleted immediately; multiple candidates prompt for clarification.

Auto-Recall (Passive)

When autoRecall is enabled, the plugin intercepts every agent turn and:

1. Takes the user's prompt
2. Searches Backboard for semantically relevant memories
3. Injects them into the system context as <relevant-memories> before the LLM sees the prompt

No tool call needed — the agent simply knows relevant context from prior conversations.

Auto-Capture (Passive)

When autoCapture is enabled, the plugin listens for completed agent turns and:

1. Extracts the last 10 messages from the conversation
2. Stores them in Backboard as a memory, tagged with source: "auto-capture" and the current session_id

This means the agent builds a memory of every conversation automatically.

CLI Commands

Inspect and search memories directly from the terminal:

# Search memories
openclaw backboard search "deployment preferences"
openclaw backboard search "favorite color" --limit 10 --scope long-term

# View stats
openclaw backboard stats

Smoke Test

Run the end-to-end smoke test to verify everything works:

./smoke_test.sh

This test:

1. Stores a memory via the agent
2. Retrieves it via semantic search
3. Confirms it in Backboard directly (CLI)
4. Checks memory stats
5. Deletes the memory via the agent
6. Confirms deletion

Architecture

┌─────────────────────────────────────────────────┐
│                  OpenClaw Gateway                │
│                                                  │
│  before_agent_start ──► auto-recall (search)     │
│  agent_end ──────────► auto-capture (store)      │
│                                                  │
│  Tools: memory_search | memory_store | ...       │
│  CLI:   openclaw backboard search | stats        │
└────────────────┬────────────────────────────────┘
                 │  HTTPS (X-API-Key)
                 ▼
┌─────────────────────────────────────────────────┐
│              Backboard.io API                    │
│                                                  │
│  /assistants  /threads  /memories  /messages     │
│                                                  │
│  Semantic search · Stateful threads              │
│  17,000+ models · Built-in RAG                   │
└─────────────────────────────────────────────────┘

Project Structure

openclaw-backboard/
├── src/
│   ├── index.ts              # Plugin definition, tools, hooks, CLI
│   ├── backboard-client.ts   # Thin HTTP client for Backboard REST API
│   └── types.ts              # TypeScript interfaces (config, API responses)
├── openclaw.plugin.json      # Plugin manifest
├── package.json
├── tsconfig.json
├── start.sh                  # Gateway launcher
└── smoke_test.sh             # End-to-end verification

License

MIT