个人Agent记忆容量存在上限,代码优于长文本规则
More memory does not keep helping personal agents, because relevant notes help u…
Agent开发者必读的实操数据,明确给出了记忆容量的Sweet Spot及代码与规则的效能对比,直接指导工程落地。
More memory does not keep helping personal agents, because relevant notes help up to a point and then extra lines start making the agent miss rules.
更多的记忆并不会持续帮助个人智能体,因为相关的笔记在达到一定限度后会有所帮助,但额外的内容反而会导致智能体忽略规则。
A personal agent can't learn every user preference through memory notes, and more notes eventually make it worse, so use code for anything it must count or track and keep memory short.
个人智能体无法通过记忆笔记学习所有用户偏好,且笔记过多最终会适得其反,因此对于必须计数或跟踪的内容应使用代码,并保持记忆简短。
Written rules work for style, like signing texts with the user's first name. For a running spending total, a stated rule still failed 44% of the time, while code that kept the total failed 0%.
书面规则适用于风格设定,例如用用户的名字签署文本。对于累计支出总额,声明式规则仍有 44% 的失败率,而用于维护总额的代码则实现了 0% 的失败率。
Memory size has a sweet spot. With Claude Haiku 4.5, violations dropped from 77% with no memory to 20% at 10 lines, then rose to about 25% with longer memories.
记忆大小存在一个最佳平衡点。在使用 Claude Haiku 4.5 时,违规率从无记忆时的 77% 降至 10 行记忆时的 20%,随后在更长记忆下回升至约 25%。
Agents that rewrite their own memory from user complaints improve early, then stall. The best methods ended near 48% violations, against 7.1% when the agent was simply told every preference.
根据用户投诉重写自身记忆的代理在初期表现改善,但随后陷入停滞。最佳方法的违规率最终接近 48%,而当代理被直接告知所有偏好时,违规率仅为 7.1%。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力