agentdb-feedback
Close the learning loop — record reward signal for an action AgentDB suggested. Use after using anything from agentdb_pattern_search / reflexion_recall / skill_search / learning_route. The bandit needs the signal to improve.
ソース情報
- リポジトリ
- ruvnet/agentdb
- ソースの最終更新活動
- 2026年5月6日 15:22
- 検出された SKILL.md の言語
- 英語
- スター
- 97
- フォーク
- 16
インストール方法
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
ソースファイルを確認
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SKILL.md を表示中
SKILL.md
ソースの指示 · 読み取り専用プレビュー- name
- agentdb-feedback
- description
- Close the learning loop — record reward signal for an action AgentDB suggested. Use after using anything from agentdb_pattern_search / reflexion_recall / skill_search / learning_route. The bandit needs the signal to improve.
# Feedback
Close the loop on a memory or routing decision so AgentDB's bandit learns.
## When to use
- **Always** after using a recall result, a routed action, or a skill suggestion. The most-undervalued part of the loop.
- After a task ends — record episode-level reward.
- When intentionally ignoring a suggestion — record *negative* reward so the bandit notices.
## API
```
agentdb_record_feedback(
id: <pattern/skill/episode/decision id>
reward: -1..1
context?: { task, outcome, latency, ... }
)
agentdb_bandit_update(
arm: <bandit arm name>
reward: -1..1
)
```
## Reward conventions
| Outcome | Reward |
|---|---|
| Used the suggestion, task succeeded | **+1.0** |
| Used the suggestion, task partial success | **+0.5** |
| Used the suggestion, didn't help | **0.0** |
| Used the suggestion, made things worse | **-0.5** |
| Ignored the suggestion (other reason) | **-0.1** (mild downweight) |
| Rejected as wrong / harmful | **-1.0** |
## Pattern: bracket every recall with feedback
```
const hits = await agentdb_pattern_search(query)
for (const h of useful(hits)) {
use(h)
await agentdb_record_feedback(h.id, +1)
}
for (const h of skipped(hits)) {
await agentdb_record_feedback(h.id, -0.1)
}
```
Without negative feedback, the bandit only sees "winners" and exploration starves.
## Don't
- Don't aggregate. Per-id feedback is what the bandit consumes; a single "task succeeded" reward attached to the task itself doesn't update individual memory weights.
- Don't fabricate reward. Honest 0.0 ("retrieved but didn't help") teaches more than dishonest +1.
- Don't only reward — *especially* record negative reward. It's the higher-information signal.
GitHubで見る