‹ 首页

recall-before-claim

@bitterbot-ai · 收录于 1 周前 · 上游提交 1 周前

Forces a memory_search before the agent sends a message containing a factual assertion that has not yet been grounded this turn. Closes the citation-rate gap from ~40% to ~90%+.

适合你,如果希望智能体回答时能主动引用已有记忆,减少编造

/ 通过 npx 安装 校验哈希
npx oh-my-skill add bitterbot-ai/bitterbot-desktop/recall-before-claim
/ 通过 bash 安装
curl -fsSL https://oh-my-skill.com/install.sh | bash -s -- bitterbot-ai/bitterbot-desktop/recall-before-claim
/ 已经装过?验证本机副本,不用重装
npx oh-my-skill verify bitterbot-ai/bitterbot-desktop/recall-before-claim
安装目标可用 --agent / --scope 或 --to 明确指定;省略时只会在唯一已存在的 agent 目录上自动选择,零命中或多命中会停止并提示。content_hash 缺失或不一致均拒装。
2425GitHub stars
~412上下文体积 · 单文件
索引托管

怎么用

商店整理自技能原文 · 版本 867a130 · 表述以原文为准
它做什么

Claude 在回复前,如果发现消息中包含未经核实的事实性断言(如“X是Y”),会先自动执行一次记忆搜索,确保引用已有记忆后再发送。

什么时候触发

当 Claude 准备发送的消息中包含事实性断言,且本轮对话中尚未使用过记忆搜索工具时触发。

装好后可以这样说
Claude 会先搜索记忆再回答。
触发记忆搜索,确保回答有依据。
技能原文 SKILL.md作者撰写 · MIT · 867a130

recall-before-claim

This skill installs a deterministic guard around the agent's outgoing messages. When the candidate message contains a factual-assertion shape ("X is Y", "A did B") and no memory tool (memory_search / deep_recall / knowledge_graph_search) has fired in the current turn, the interceptor requires a memory_search call against the assertion's subject before the message goes out.

Why this exists

Bitterbot has a strong memory system, but it only helps the user if the agent actually consults it before making claims. In practice the agent skips memory recall a meaningful fraction of the time — relying on its in-context knowledge instead. This produces ungrounded assertions that the user has no easy way to spot.

The interceptor closes that loop without depending on the LLM remembering to do it.

What you'll see

When this fires, the agent will run a memory_search first, integrate the result, and then send. You may notice slightly longer responses to factual questions; you should notice that the agent stops making confidently wrong statements about things it actually has memory of.

How it decides not to fire
  • Opinion shapes ("I think", "maybe", "in my opinion") are skipped.
  • Questions are skipped.
  • If a memory tool already fired in the last ~30 seconds, the assertion is considered grounded.
Implementation

Built-in interceptor recall-before-claim:default lives in src/agents/skills/builtin-interceptors/recall-before-claim.ts. Fires at most 8 times per session.

按 MIT 许可原样转载,未经改动 · 在 GitHub 查看 →

评论

登录即可评论;带「已验证安装」的,是发布者名下有本店的安装或持有记录。