[ENH] Memory decay and relevance-based eviction in EMC

[OppaAI]

Summary

EMC currently stores everything forever with no decay. Over time this creates "memory clutter" — irrelevant old episodes compete equally with important recent ones during search.

Current Behavior

All episodes have equal weight regardless of age or relevance. A trivial exchange from 2 years ago ranks the same as an important conversation from yesterday.

Proposed Enhancement

Add relevance-based decay to EMC search scoring:

def _score_episode(self, episode: dict, query_vec: list, now: datetime) -> float:
    # Semantic similarity
    semantic = _cosine(query_vec, json.loads(episode["embedding"]))
    
    # Time decay — older episodes score lower
    age_days = (now - datetime.fromisoformat(episode["date"])).days
    decay = math.exp(-age_days / DECAY_HALF_LIFE)  # exponential decay
    
    return semantic * decay

Options:

• Exponential decay — gradual fade, important things still surface if relevant
• TTL-based — hard expiry after N days (too aggressive)
• Access-based — episodes accessed frequently stay fresh (like human memory)
• Importance threshold — only decay below a relevance threshold

Impact

• Recent important memories rank higher naturally
• Old trivial memories fade gracefully
• More human-like memory behavior
• Reduces "memory clutter" over years of use

Notes

• Decay should be gentle — GRACE remembers important things forever
• DECAY_HALF_LIFE should be configurable in mcc.yaml
• Implement in M2 alongside SMC — the 11pm reflection is a natural point to apply decay scoring
• Do NOT implement hard deletion — 1TB NVMe means storage is free, keep everything