enhanced-memory
jameseball 제작
Search & Research요약
Enhanced memory search with hybrid vector+keyword scoring, temporal routing, filepath scoring, adaptive weighting.
설명
이 스킬이 하는 일
A drop-in replacement for OpenClaw's default memory search. Uses a 4-signal fusion pipeline combining vector similarity, keyword matching, header matching, and filepath scoring to retrieve memory chunks. Also includes salience scoring and a knowledge graph cross-reference builder.
Achieves 0.782 MRR versus approximately 0.45 for baseline vector-only search by fusing four signals and applying adaptive weighting when keyword overlap is low.
사용 시점
- Finding relevant past context when vector-only search returns weak results
- Surfacing stale memory files during agent heartbeat cycles
- Building cross-reference links between related memory chunks
- Querying memories with date references like 'last Monday' or 'Feb 8'
- Indexing updated memory files after significant workspace changes
예시 워크플로
User asks: 'What did I decide about the auth system last week?'
- 1
Detects temporal reference 'last week' and applies 3x boost to date-matching files
- 2
Embeds the query using nomic-embed-text via Ollama
- 3
Scores all memory chunks across vector similarity, keyword overlap, header match, and filepath
- 4
Checks if top score exceeds PRF threshold; if not, expands query terms and re-scores
- 5
Returns ranked results with path, score, and snippet
Ranked memory chunks referencing auth system decisions from the relevant timeframe
요구 사항
- Ollama installed and running locally on port 11434
- nomic-embed-text model pulled via Ollama
설치
clawhub install jameseball/enhanced-memoryCLI 명령을 실행하거나, 프롬프트를 Claude Code, Cursor, Codex 등 사용하는 코딩 에이전트에 붙여넣으세요.