Imported from huathanhnam95-del/demoproject (.claude/skills/agent-memory-ultimate/SKILL.md). Install upstream with npx skills add huathanhnam95-del/demoproject --skill agent-memory-ultimate. Copyright stays with the author.

Agent Memory Ultimate

Enable persistent semantic storage and structured long-term reasoning for AI agents.

Use this skill when

• Implementing long-term context across conversation sessions.
• Using MRL (Matryoshka Representation Learning) for efficient vector searches.
• Structuring reasoning via "Sequential Thinking" (scratchpads).
• Building Knowledge Items (KIs) from unstructured data.

Logicware Components

1. Sequential Thinking

Store intermediate thoughts to maintain a clear "chain of command" and prevent logic drift.

• Protocol: Thought -> Observation -> Refinement -> Action.

2. MRL Indexer (k3_mrl_indexer.py)

• Model: text-embedding-004
• Batching: Process in chunks of 5-10 for rate-limit stability.
• Search: Rank snippets by cosine similarity and filter by confidence thresholds.

Implementation Workflow

1. Index: Scan target directories and generate .pkl semantic indices.
2. Retrieve: Query index with user prompt to find relevant past contexts.
3. Reason: Apply Sequential Thinking to weave retrieved data into the current plan.
4. Update: Persist new findings back to the Knowledge Base.

Metrics

• Recall@K: Measure how often the correct context is in the top K results.
• Coherence: Ensure reasoning steps follow logically from retrieved memory.