| name | memory-crystallization |
| description | Transforms raw session logs and temporary memories into high-density, structured semantic knowledge. |
| version | 1.0.0 |
| author | Hermes Agent (Self-Evolved) |
| tags | ["memory","optimization","continuity"] |
Purpose
This skill was self-evolved from the GitHub fork upgrade process to enhance the agent's capabilities.
Memory Crystallization
This skill implements the "Crystallization" loop from agentmemory. It prevents context window bloat and memory decay by refining raw data into durable facts.
The Crystallization Pipeline
- Ingestion: Scan current session logs and
~/.hermes/memory/ for raw, unrefined observations.
- Reflection: Use the internal
reflect loop to identify recurring patterns, critical technical facts, and user preferences.
- Deduplication: Compare new insights with existing memories to prevent redundant entries.
- Crystallization: Convert a "finding" into a declarative fact.
- Raw: "User mentioned he prefers using the local gemma4 model over cloud because of rate limits."
- Crystallized: "User prefers local gemma4 over cloud due to rate limiting constraints."
- Commit: Save to persistent memory with a timestamp and context tag.
Execution Workflow
- Scan: Run
session_search for the last 5 sessions.
- Filter: Extract all "Technical Decisions," "User Preferences," and "Project Architecture" facts.
- Refine: For each fact:
- Remove conversational filler.
- Verify against existing memory.
- Normalize terminology (e.g., "local model" -> "local gemma4").
- Persist: Use
memory(action='add', target='memory', content='...').
Pitfalls
- Over-Compression: Losing the "Why" behind a decision. Always preserve the reasoning if it's non-obvious.
- Stale Truths: Crystallizing a fact that has since been changed. Always check for contradictions before committing.
Verification
- Run
memory list to ensure a concise, non-redundant set of facts.
- Verify that a
session_search on a specific topic now yields a "Crystallized" result rather than a raw transcript.
