{
  "version": "1.0",
  "description": "Attention-weighted context window KV cache weight assignments",
  "weights": {
    "system_prompt": {
      "category": "system",
      "weight": 2.8,
      "compression": "none",
      "rationale": "High priority, always attended"
    },
    "active_task_context": {
      "category": "system",
      "weight": 2.5,
      "compression": "none",
      "rationale": "Current work, needs focus"
    },
    "recent_chat_last_5_turns": {
      "category": "temporal",
      "weight": 1.5,
      "compression": "light",
      "decay_rate": 0.1,
      "rationale": "Temporally relevant"
    },
    "older_chat_history": {
      "category": "temporal",
      "weight": 0.65,
      "compression": "medium",
      "decay_rate": 0.3,
      "rationale": "Decaying relevance"
    },
    "memory_cards_topic_matched": {
      "category": "memory",
      "weight": 0.7,
      "compression": "light",
      "rationale": "Useful but not urgent"
    },
    "memory_cards_topic_unmatched": {
      "category": "memory",
      "weight": 0.4,
      "compression": "heavy",
      "rationale": "Low signal for current task"
    },
    "voice_transcriptions": {
      "category": "sensory",
      "weight": 0.4,
      "compression": "heavy",
      "rationale": "Low signal unless topic-relevant"
    },
    "reflection_notes": {
      "category": "structural",
      "weight": 0.5,
      "compression": "medium",
      "rationale": "Periodic, not continuous"
    },
    "scenario_state_metadata": {
      "category": "structural",
      "weight": 0.6,
      "compression": "medium",
      "rationale": "Structural, not content-heavy"
    }
  },
  "compression_threshold": 0.4,
  "expected_token_reduction_pct": 42,
  "research_backing": [
    "Chen et al., Distilling Task-Specific Knowledge from BERT into Transformer (2019)",
    "Child et al., Generating Long Sequences with Sparse Transformers (2019)"
  ]
}