Auto-recall: inject relevant context before agent turns

[bm-clawd]

Context

Currently, agents have to explicitly call bm_search or memory_search to find relevant knowledge. The plugin should proactively inject relevant context from the knowledge graph before the agent starts its turn.

This is similar to how memory-lancedb works — automatic context injection based on the user's message.

What this enables

• Agent automatically "remembers" relevant context without being told to search
• Reduces tool calls (no explicit search step needed)
• Better continuity across sessions — agent wakes up with relevant context loaded
• Knowledge graph becomes invisible infrastructure rather than an explicit tool

Implementation

• Hook into agent_start or message preprocessing
• Extract key terms/intent from the user's message
• Search the knowledge graph for relevant notes
• Inject top N results as system context (similar to how OpenClaw injects workspace files)
• Configurable: max results, relevance threshold, max tokens
• Should be fast — add latency budget config

Config example

{
  "basic-memory": {
    "config": {
      "autoRecall": {
        "enabled": true,
        "maxResults": 3,
        "maxTokens": 2000,
        "minScore": 0.5
      }
    }
  }
}

Notes

• Listed as post-1.0 in TODO but this is a high-value differentiator
• Need to be careful about latency — search must be fast
• Consider caching recent searches to avoid redundant lookups