ReMe memory for OpenClaw

中文说明

ReMe gives OpenClaw file-native long-term memory while keeping durable memory in a workspace you own. The integration uses OpenClaw's public plugin SDK, memory slot, authenticated Control UI, service lifecycle, conversation hooks, and tool protocol.

Capabilities

• reme_search performs an explicit, source-backed memory search.
• before_prompt_build recalls relevant memory before a conversational root-agent turn.
• agent_end captures the original user/assistant pair in serialized background batches.
• session_end flushes a session boundary; service shutdown drains all work within a bounded budget.
• Auto Dream runs once per configured day and consolidates daily notes into durable digest knowledge.
• The ReMe Memory sidebar tab shows health, capture settings, Auto Dream state, and component diagnostics.

Recalled text is wrapped in <reme-context>, marked as untrusted historical data, and escaped so memory content cannot close the wrapper. Subagent, cron, heartbeat, memory, and overflow runs are excluded by default.

Compatibility and prerequisites

• OpenClaw 2026.9.3 or newer.
• Node.js 24.16.0+ on Node 24, or 26.1.0+ on Node 26.
• Python 3.11+ and a ReMe HTTP service exposing search, auto_memory, auto_dream, health_check, and status.
• A model configured for ReMe's automatic-memory and Auto Dream jobs.

The plugin never receives an LLM API key. Model credentials belong to ReMe/OpenClaw configuration and must not be put in this package, a screenshot, or a committed config file.

1. Start ReMe

Install ReMe, choose a user-owned workspace, and bind the service to loopback. Port 3458 is useful when another ReMe instance already uses the default port.

pip install "reme-ai[core]"
reme start \
  workspace_dir=/absolute/path/to/reme-workspace \
  service.host=127.0.0.1 \
  service.port=3458

Verify the service without exposing configuration or credentials:

curl -fsS -X POST http://127.0.0.1:3458/health_check \
  -H 'Content-Type: application/json' -d '{}'

ReMe HTTP does not add API-key authentication. Keep it on loopback or place it behind a trusted authenticated proxy.

2. Install the plugin

From a released package:

openclaw plugins install clawhub:@agentscope-ai/reme-openclaw-plugin

From this repository, build and install the exact archive that was tested:

cd /path/to/ReMe/integrations/openclaw
npm ci
npm run build
npm pack
openclaw plugins install --force ./agentscope-ai-reme-openclaw-plugin-0.1.0.tgz

Restart the Gateway after installation. In Settings → Plugins, search for ReMe; it should be enabled, categorized as Memory, and expose reme_search.

The plugin details also show the required conversation grant. No secret fields are part of the plugin schema.

3. Configure OpenClaw

Add the following to the effective OpenClaw config, then restart the Gateway:

{
  "plugins": {
    "slots": { "memory": "reme" },
    "entries": {
      "reme": {
        "enabled": true,
        "hooks": { "allowConversationAccess": true },
        "config": {
          "endpoint": "http://127.0.0.1:3458",
          "language": "en",
          "autoRecall": true,
          "autoMemoryEnabled": true,
          "autoMemoryInterval": 5,
          "autoDreamEnabled": true,
          "dreamCron": "0 23 * * *",
          "timezone": "Asia/Shanghai"
        }
      }
    }
  }
}

Both settings outside config are required: plugins.slots.memory makes ReMe the active memory provider, while allowConversationAccess permits the non-bundled plugin to inspect completed conversation turns. Without the latter, explicit search may work while automatic capture does not.

Validate and inspect the effective runtime:

openclaw config validate --json
openclaw plugins inspect reme --runtime --json
openclaw gateway status

Runtime inspection should report status: loaded, memorySlotSelected: true, reme_search, three typed hooks, one service, and one authenticated HTTP route. Current plugins validate targets authoring-metadata-only tool/feature plugins; use runtime inspection for this mixed lifecycle plugin.

Configuration reference

| Option | Default | Meaning | | --------------------- | ----------------------- | ------------------------------------------------- | | endpoint | http://127.0.0.1:2333 | ReMe HTTP service URL | | language | en | Memory guidance language: en or zh | | autoRecall | true | Recall before conversational root-agent turns | | searchLimit | 5 | Maximum returned search results | | recallMinScore | 0 | Minimum score for automatic recall | | autoMemoryEnabled | true | Capture completed user/assistant turns | | autoMemoryInterval | 5 | Completed turns per capture batch | | autoDreamEnabled | true | Enable scheduled consolidation | | dreamCron | 0 23 * * * | Daily schedule (minute hour * * *) | | dreamHint | empty | Optional guidance passed to auto_dream | | rootAgentsOnly | true | Exclude subagents and non-conversational triggers | | timezone | Asia/Shanghai | IANA timezone used for batching and scheduling | | requestTimeoutMs | 10000 | Recall, search, and status timeout | | backgroundTimeoutMs | 3600000 | Automatic-memory and Auto Dream timeout | | shutdownTimeoutMs | 5000 | Best-effort Gateway shutdown drain budget |

Failed capture batches are retained in memory for retry. Durable state is written only by ReMe into its configured workspace; plugin queues and diagnostics are process-local and rebuildable.

4. Use and verify

Status frontend

Open ReMe Memory in the sidebar. /plugins/reme/status/ is protected by Gateway auth. The page is server-rendered because OpenClaw embeds external plugin tabs in a script-free sandbox. It shows the endpoint and component metrics, but not model credentials or conversation content.

Explicit search

Ask: Use reme_search to look up Project Lighthouse. The transcript should show a ReMe Search tool invocation and the answer should cite workspace-relative memory paths.

Automatic recall

Start a new session and ask a question existing memory can answer, adding Without calling tools. A correct answer with no tool card proves that before_prompt_build injected recalled context.

Automatic memory across sessions

For a quick test, temporarily set autoMemoryInterval to 1. Tell OpenClaw a unique, synthetic fact, wait for the turn and ReMe background job to finish, then ask for it in a different session without calling tools. Also confirm that a new Markdown note exists beneath the configured ReMe workspace.

Restore a larger interval after testing to reduce model calls.

Auto Dream

Wait for dreamCron, then reload ReMe Memory → Memory Consolidation, verify Last result: completed, and inspect the ReMe digest update. The plugin status route is read-only; memory consolidation runs only through the configured scheduler.

Troubleshooting

• Plugin is loaded but automatic recall/capture is absent: check the memory slot and hooks.allowConversationAccess; restart after changing either.
• ReMe Memory says Offline: call ReMe health_check, verify the port, and ensure both processes can reach the same loopback/network namespace.
• Explicit search works but capture does not: verify autoMemoryEnabled, use interval 1 for diagnosis, and check ReMe logs for /auto_memory requests.
• Capture waits after a short conversation: the default is five completed turns. Session end and Gateway shutdown also attempt a bounded flush.
• Auto Dream never runs: the accepted cron form is daily minute hour * * *; confirm the IANA timezone and next run shown in the status tab.
• Gateway rejects the plugin: use a supported Node/OpenClaw version, rebuild, then inspect runtime diagnostics.
• Status tab is blank after an upgrade: confirm runtime inspection reports one authenticated HTTP route and restart the Gateway. Do not relax iframe sandboxing; this integration is designed for it.

Development checks

cd integrations/openclaw
npm ci
npm run format:check
npm run lint
npm run typecheck
npm test
npm run test:package

Tests mock network boundaries. Real E2E work should use a temporary ReMe workspace and synthetic data; never commit .env, runtime sessions, indexes, logs, caches, packed archives, or generated test memory.

Source and license

ReMe is developed at agentscope-ai/ReMe and released under Apache-2.0.