Local notes help one coding agent, but a shared memory is what helps the team

说明与用途

本地笔记帮助单个代理，但共享内存显著提升团队效率，降低修复成本和验证修复时间。

问题

Distributed coding agents need shared knowledge to efficiently fix bugs that cross team members and repositories.

结果

Local memory boosts fix rates when the same agent repeats a bug, but shared BAB significantly improves teammate fix rates and lowers team cost per fix across models.

工作流

• Simulate three-agent teams with isolated sandboxes
• Define three wall scenarios (Paylane, Gatekeeper, geo-api) across two repos
• Run each agent to fix a bug in the first repo
• Freeze memory and run a related bug in the second repo
• Compare no memory, local-only memory, and local+shared BAB setups
• Measure cost per verified fix and time to fix across 378 graded sessions
• Analyze results per model and team level

证据

Local memory is excellent when the same agent hits the same wall

Local notes stop helping the moment the bug lands on a teammate's machine

In this distributed round, with the shared BAB in place, the teammate fix rate was 100% for Sonnet and Opus and 96% for Haiku

Team cost per delivered fix, including first-solver sessions, for no memory, local-only and shared memory across three models

来源