@cortexmem/cortex-bridge

Bridge plugin that connects OpenClaw agents to Cortex memory service.

Uses OpenClaw's standard register(api) plugin interface.

Install

openclaw plugins install @cortexmem/cortex-bridge

Configure

Plugin config via OpenClaw settings or environment variables:

| Config Key | Env Variable | Default | Description | |------------|-------------|---------|-------------| | cortexUrl | CORTEX_URL | http://localhost:21100 | Cortex server URL | | agentId | — | openclaw | Agent identifier for memory isolation | | debug | CORTEX_DEBUG | false | Enable debug logging |

Tools

These tools are always available and work reliably:

| Tool | Description | |------|-------------| | cortex_recall | Search long-term memory for relevant past conversations, facts, and preferences | | cortex_remember | Store a fact, preference, or decision (supports category: fact, preference, skill, identity, etc.) | | cortex_ingest | Send a conversation pair for automatic LLM memory extraction | | cortex_health | Check if the Cortex memory server is reachable (optional) |

Slash Command

• /cortex-status — Quick check if Cortex server is online

Hooks

The plugin registers lifecycle hooks for automatic memory management:

| Hook | Status | Description | |------|--------|-------------| | before_agent_start | Working | Recalls relevant memories and injects as context before each response | | agent_end | Not working | Should auto-ingest conversations after each response (see Known Issues) | | before_compaction | Best-effort | Emergency flush before context compression |

Known Issues

agent_end hook not firing in streaming mode

Status: Upstream bug — openclaw/openclaw#21863

In streaming mode (used by Telegram and other gateway channels), the agent_end hook is not dispatched to plugins. The handleAgentEnd() function in OpenClaw's streaming event handler does not call hookRunner.runAgentEnd().

This means automatic conversation ingestion does not work in streaming mode. Memory recall (before_agent_start) works correctly.

Workarounds:

1. Use cortex_ingest tool — Instruct your Agent (via system prompt) to call cortex_ingest after meaningful conversations. Example system prompt addition:

   After each conversation, use the cortex_ingest tool to save the exchange
   for long-term memory. Pass the user's message and your response.

2. Use non-streaming mode — If your setup supports it, use a non-streaming channel where agent_end fires correctly.

3. Use cortex_remember tool — For specific facts or preferences, the Agent can call cortex_remember directly during conversation.

License

MIT