openclaw-memorysync

MemorySync memory backend for OpenClaw — automatic recall and capture on every interaction, backed by MemorySync. Takes OpenClaw's exclusive memory slot, so your assistant remembers you across every channel it lives on: WhatsApp, Telegram, Discord, Signal, and the rest.

Install

openclaw plugins install npm:openclaw-memorysync

Then wire it up in ~/.openclaw/openclaw.json (get a key at app.memorysync.io):

{
  plugins: {
    slots: { memory: "openclaw-memorysync" },
    entries: {
      "openclaw-memorysync": {
        enabled: true,
        hooks: {
          allowPromptInjection: true,      // recall injection
          allowConversationAccess: true    // capture
        },
        config: { apiKey: "${MEMORYSYNC_API_KEY}" }
      }
    }
  }
}

openclaw gateway restart

Ask /memorysync-status in any chat — it diagnoses the key, the slot, both permission gates, the allowlist, and live connectivity, and prints the exact config to paste for anything missing.

What runs when

| Moment | What happens | | --- | --- | | Before each reply | Memories relevant to your message are recalled and injected — inside a hard time budget, so a slow network can never delay a chat. | | After each reply | The exchange persists verbatim with content-hash idempotency seeds (retries converge on one stored row; the plugin's own injected context is stripped first). | | Session start | Tenant cache warm-up, fire-and-forget. | | Session end | Cache trim only — never network (OpenClaw's 2-second drain budget is respected by design). | | Any failure — no key, network down, monthly quota exhausted | Silent skip. A raw error never reaches your chat; the worst case is a memoryless reply. |

Tools

memory_search, memory_add, memory_get, memory_list, memory_update, memory_delete (single-id only — there is deliberately no delete-all), and memory_status (the doctor). Failures answer with friendly text, never stack traces. Credential-looking text is refused client-side before it can be stored.

Skills / commands

• /remember <fact> — save a durable fact (dispatches straight to memory_add)
• /recall <query> — search memory (straight to memory_search)
• /memorysync-status — the setup doctor
• A model-facing skill teaches the agent when to search, what to save, and to treat recalled text as background data — never instructions.

Configuration

| Key / env | Default | Meaning | | --- | --- | --- | | config.apiKey / MEMORYSYNC_API_KEY | — | Required for memory. Without it everything is a silent no-op. | | config.userId / MEMORYSYNC_USER_ID | OS username | Memory identity. | | config.baseUrl / MEMORYSYNC_BASE_URL | https://api.memorysync.io | Self-hosted / regional override. | | config.autoRecall | true | Inject memories before each turn. | | config.autoCapture | true | Persist each exchange. | | config.topK | 8 | Memories per recall. | | config.recallTimeoutMs | 6000 | Recall gives up after this and the turn proceeds memoryless. | | MEMORYSYNC_DISABLE=1 | — | Switch everything off without uninstalling. |

Multi-agent setups get isolated transcripts automatically (openclaw::<agentId> scopes) while sharing the same user memories.

MCP alternative

Prefer plain MCP tools without the memory slot? MemorySync's hosted MCP server works in OpenClaw directly:

openclaw mcp add memorysync --url https://mcp.memorysync.io/mcp --transport streamable-http

Docs

Full guide: docs.memorysync.io/guides/openclaw