Repository navigation
DSH|tmem|Zero-Token Deterministic Engineering Memory & Session Digest Plugin | 零 Token 消耗的仓库级确定性工程记忆与历史会话摘要插件 #13464
Replies: 2 comments
|
目前项目没有对真实轨迹做个太多复测验证,非常希望大家可以结合实际场景进行验证反馈意见。 对比LLM总结的技能,memory,在总结摘要后变成了玄学,高频场景下缺少验证。 |
|
So far the project hasn't put much effort into re-testing and validating against real trajectories, and we'd really like everyone to validate against actual usage scenarios and share feedback. The core idea is to optimize effectively for repetitive, well-defined requests — the scenarios are quite clear-cut, so this approach isn't meant for exploratory tasks or tasks that require continuous iteration. Compare that with LLM-distilled skills and memory: once they're condensed into a summary, they tend to turn into something quasi-mystical, and there's no validation in high-frequency scenarios. So let's try a different angle: treat real tool invocations as the factual dimension. At the same time, past trajectories were themselves purchased with real tokens, so there's no reason to waste them. Memory isn't free either. The goal is to strike the best balance we can between the "memory tax" and actual effectiveness. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
一句话总结:历史任务工具调度(含思考回复),在重复类型的任务中本身是真实的记忆锚点,以此为核心构筑记忆系统,拒绝语义总结。
All reactions