agentdb-hierarchical-store
Store memories in tier-aware hierarchical memory — working / short-term / long-term — and recall with tier filters. Use for working-set context that should fade, vs facts that should persist, vs patterns that should be searchable forever.
소스 정보
- 저장소
- ruvnet/agentdb
- 최근 소스 활동
- 2026년 5월 6일 15:22
- 감지된 SKILL.md 언어
- 영어
- 스타
- 89
- 포크
- 14
설치 방법
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
소스 파일 검토
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
SKILL.md 표시 중
SKILL.md
소스 지침 · 읽기 전용 미리보기- name
- agentdb-hierarchical-store
- description
- Store memories in tier-aware hierarchical memory — working / short-term / long-term — and recall with tier filters. Use for working-set context that should fade, vs facts that should persist, vs patterns that should be searchable forever.
# Hierarchical Store
Three-tier memory matching how humans actually use memory. Each tier has different retention, retrieval cost, and query semantics.
## Tiers
| Tier | Retention | Use for |
|---|---|---|
| **working** | minutes (auto-expires) | The current task's scratch — files open, hypotheses, in-flight decisions |
| **short** | hours-days | Recent session context — what was decided this morning |
| **long** | indefinite (until pruned) | Cross-session facts, patterns, lessons |
## API
```
agentdb_hierarchical_store(
key: <namespace key> // e.g. 'task:auth-refactor:hypothesis'
tier: 'working' | 'short' | 'long'
value: <stringified content>
ttl?: <seconds> // overrides tier default
metadata?: { topic, project, ... }
)
agentdb_hierarchical_recall(
query: <semantic or exact key>
tier?: <filter> // omit to search all tiers
k?: 5
)
agentdb_hierarchical_delete(
key: <namespace key>
tier?: <filter>
)
```
The delete tool was added in agentdb 3.0.0-alpha.13. Before that, there was no first-class way to remove an entry — re-indexing workflows like ruflo's `/adr-index` couldn't purge stale ADRs. Now they can.
## Recall flow
1. Query `working` first — it's smallest, freshest, highest signal for "what am I doing right now."
2. Fall back to `short` if working returns nothing.
3. Search `long` last for established knowledge.
4. Or search all three with `tier=undefined` and let the bandit weight by tier × similarity.
## Don't
- Don't put long-term facts in `working` — they expire and you lose them.
- Don't put session scratch in `long` — it pollutes searches forever.
- Don't forget to set TTL when overriding tier defaults; the engine won't infer it.
GitHub에서 보기