@dmemo/openclaw-plugin

OpenClaw host adapter for dMemo (T3.3). Forked from @mem0/openclaw-mem0's lifecycle shape (isolation scoping, deterministic before_prompt_build recall, agent_end capture, dream-consolidation gate) with the OSS Memory init's OpenAI-by-default embedder/LLM swapped for @dmemo/core's local embedder + DmemoSession's 0G Storage flush/restore. All mem0-Platform-only surfaces are stripped, not ported.

Hard requirement, structural not conventional: the shipped defaults can never call OpenAI. Embedding is always local (@dmemo/core's default embedder); capture always passes infer: false to memory.add() unconditionally — this plugin has no config knob that could produce a real OpenAI call.

Install

openclaw plugins install @dmemo/openclaw-plugin

Then register it in your OpenClaw config (standard plugins.entries.<id>.config block, same shape every OpenClaw memory backend uses):

{
  "plugins": {
    "slots": { "memory": "dmemo" },
    "entries": {
      "dmemo": {
        "config": {
          "privateKey": "${DMEMO_PRIVATE_KEY}",
          "scope": "default",
          "network": "testnet",
          "recall": { "strategy": "smart", "topK": 5, "timeoutMs": 10000 },
          "dream": { "enabled": true, "minHours": 24, "minSessions": 5, "minMemories": 20 }
        }
      }
    }
  }
}

plugins.slots.memory is exclusive — registering "dmemo" here takes over recall/capture from whatever memory plugin (if any) owned the slot before.

DMEMO_PRIVATE_KEY — the easiest way to get this value is npx @dmemo/cli setup, which writes it to ~/.dmemo/config.json; export it into the environment OpenClaw runs in, or paste the value directly (not recommended — prefer the env-var interpolation form above).

Config reference (openclaw.plugin.json)

| Key | Type | Default | Notes | |---|---|---|---| | privateKey | string | — | 0x-prefixed hex. Required — no config, no memory (fails open, never throws). | | scope | string | "default" | Base mem0 user_id. Per-agent sessions auto-namespace as <scope>:agent:<agentId>. | | network | "testnet" \| "mainnet" | "testnet" | | | recall.strategy | "always" \| "smart" \| "manual" | "smart" | | | recall.topK | number | 5 | | | recall.timeoutMs | number | 10000 | Raised from a generic default to clear T0.1's measured cold-restore time with headroom. | | dream.enabled | boolean | true | Tags dream mutations source: "dream"; flushes a dream burst as one delta batch. | | dream.minHours / minSessions / minMemories | number | 24 / 5 / 20 | Dream-consolidation gate thresholds. |

Known limitation

infer is hardcoded to false unconditionally in this build — there is no config knob that flips it, and setting one in a future version would need @dmemo/core's session-open seam to grow a real LLM slot wired to the 0G Router first. See the infer field's doc comment in src/config.ts for why: passing infer: true through today would make a real network call to a placeholder OpenAI LLM config — exactly the "shipped defaults call OpenAI" failure mode this plugin exists to prevent.

Inference routing (optional, separate from the memory leg)

To route OpenClaw's own chat inference through the 0G Compute Router:

{
  "models": {
    "providers": {
      "zg": {
        "baseUrl": "https://router-api-testnet.integratenetwork.work/v1",
        "api": "openai-completions",
        "apiKey": "${ZEROG_API_KEY}"
      }
    }
  },
  "agents": { "defaults": { "model": { "primary": "zg/qwen2.5-omni" } } }
}

Testnet model catalog drift (verified live 2026-07-25): the spec-era qwen/qwen2.5-omni-7b (TeeML) id no longer exists; the current testnet chat model is qwen2.5-omni (TeeTLS, not TeeML — see docs/disclosure.md). Re-check GET {baseUrl}/models before pinning a model id in production. Mainnet Router: https://router-api.0g.ai/v1.

Development

pnpm run build   # tsc -b
pnpm test        # node --test against dist/*.test.js (17/17, mocked, no network)