hierarchical-memory-index

name	hierarchical-memory-index
description	Implements a multi-tier memory system (Immediate -> Short-Term -> Long-Term/Crystallized) to optimize recall and minimize context bloat.

Hierarchical Memory Indexing (LCM Implementation)

This skill replaces linear memory files with a three-tier retrieval system inspired by the hermes-lcm architecture.

The Memory Tier System

1. L1: Working Memory (Context Window)

   • Current session state, immediate goals.
   • Action: Managed via TASK_STATE.md and active context.

2. L2: Short-Term Semantic Cache (Daily/Weekly Notes)

   • High-fidelity logs of recent events.
   • Action: Managed via memory/YYYY-MM-DD.md.

3. L3: Long-Term Crystallized Core (Global Truths)

   • De-duplicated, validated facts.
   • Action: Managed via memory.md and USER.md.

The Indexing Loop

• Promotion: When a fact in L2 is referenced 3+ times across different sessions -> Promote to L3 (Crystallize).
• Demotion/Eviction: When L2 notes exceed a size threshold -> Summarize into L3 and archive raw log.
• Cross-Reference: Every L3 fact should link back to the L2 session date that generated it for provenance.

Execution Workflow

1. Audit: Scan L2 files for recurring patterns.
2. Synthesize: Use the memory-crystallization skill to create a durable fact.
3. Index: Map the fact to a specific "Domain" (e.g., #architecture, #user-pref, #bmo-stack).
4. Verify: Ensure no contradictions exist between the new L3 fact and existing truths.

Verification

• Run session_search using a domain tag (e.g., "search #architecture") and verify the result is a distilled L3 fact.