Smarter auto-capture: LLM-based importance scoring and summarization

[hd719]

Problem

Auto-capture uses regex pattern matching (MEMORY_TRIGGERS) which is brittle:

• Misses important context that doesn't match patterns (e.g., 'My dog is named Rex')
• Captures noise that happens to match (e.g., 'I like this color' in a throwaway comment)
• No summarization — stores raw user text which may be verbose or contain irrelevant detail
• Category detection is also regex-based and frequently miscategorizes

Proposal

Add an optional LLM-based capture pipeline:

1. Importance scoring: After regex pre-filter, send candidate messages to a lightweight LLM call to score importance (0-1) and decide whether to store
2. Summarization: Distill the raw message into a clean, concise memory statement before embedding
3. Better categorization: Let the LLM classify into categories with higher accuracy
4. Context-aware capture: Include the assistant's response for context (currently only user messages are evaluated)

Config

{
  "autoCapture": true,
  "captureMode": "llm",  // or "regex" (default, current behavior)
  "captureModel": "gpt-4o-mini",
  "captureMinImportance": 0.6
}

Tradeoffs

• Adds API cost per conversation (mitigate with batch/cheap models like gpt-4o-mini or Gemini Flash)
• Regex mode stays as default for zero-cost operation
• LLM mode would be opt-in