GitHarness 用 Git 式提交与分支管理智能体记忆
有人把智能体记忆做成了 Git 那样的提交和分支,用户改需求时只重做变化的部分,30 个测试全赢,编码场景省了 73.6% 的 token。
GitHarness 把智能体的记忆当作 Git 仓库来管理:每条需求连同对应的工作成果存成 commit,需求变更时从最近的有效版本开分支。这样智能体可以丢弃过时的需求,保留仍然有效的部分,避免把陈旧工作带到后续步骤。实验覆盖 30 种测试设置,全部优于普通的续写式延续。其中 1 个编程任务设置里,token 消耗减少 73.6%,得分反而更高。
Treating agent memory like a Git repo, with commits and branches, lets agents drop outdated requirements and keep the work that is still valid.
Agents handle mid-task changes from users better when they save versions of their work and resume from the right one, so build that versioning into any agent that takes feedback.
Users rarely give a full spec upfront. They add, fix, or drop requirements as results come in. Most agents then carry stale work forward or redo everything.
GitHarness stores each requirement with its matching work, like Git commits, and branches from the closest valid version. It beat plain continuation in all 30 tested settings. In 1 setup, it used 73.6% fewer tokens on coding tasks while scoring higher.
When a user changes 1 detail, go back to the version that still fits and redo only what changed.