Configurable similarity thresholds and recall tuning

[hd719]

Problem

Key thresholds are hardcoded with no way to tune:

• Duplicate detection: 0.95 (in memory_store)
• Auto-recall minimum similarity: 0.3 (in before_agent_start)
• Recall result count: 3 (in before_agent_start)
• Forget match threshold: 0.7 / 0.9 (in memory_forget)
• Tool recall minimum: 0.1 (in memory_recall tool) vs 0.5 (in MemoryDB.search default)

The 0.1 vs 0.5 mismatch between the tool call and the DB default is also a subtle bug — the tool passes 0.1 but the DB method signature defaults to 0.5, so behavior depends on which path is taken.

Proposal

Add config options:

{
  "thresholds": {
    "duplicateMin": 0.95,
    "recallMin": 0.3,
    "recallLimit": 3,
    "forgetAutoDeleteMin": 0.9
  }
}

Different use cases need different tuning:

• High-volume assistants may want stricter recall (0.5+) to reduce noise
• Personal assistants may want looser recall (0.2) to surface more context
• Duplicate threshold may need loosening (0.85) if users rephrase often