An Apple research team found that coding agents performed better when they retained selected reusable facts instead of carrying complete conversation histories between sessions. The proposed memory stores task specifications, data schemas, tool configurations and output constraints while dropping session-specific reasoning that may become stale.
In three enterprise deployment scenarios, the system completed 96% of tasks, compared with 79% for agents without persistent memory and 71% when the full history was preserved. The researchers say old reasoning can bias later work, while carefully selected context gives an agent the facts it needs without filling its context window. Shared workspaces can transfer that memory between users under role-based access controls.
The implementation also keeps generated artifacts in Git, isolates drafts and lets users restore earlier versions. A separate refresh mechanism updates recurring data without invoking the language model again; the paper reports a 14-fold reduction in task time for those updates and a 97-fold reduction in per-invocation token cost when summaries replace raw data. These are research results from a particular collaborative platform, not a guarantee for every agent. Still, they offer a concrete warning against treating an ever-growing transcript as an adequate memory system.