Persistent memory enables language‑model agents to improve prompts and skills without altering model weights. We show that aligning the retrieval scope with the certification scope allows such edits to support reliable repeated adaptation across recurring task families. Experiments were conducted with frozen‑model agents on ProcStream‑RSI, a 12‑round code‑repair stream, using Orthogonal Regression Control (ORC) – an execution‑grounded gate for persistent skill edits.
In an intervention where proposals and gate decisions were fixed, retrieving each accepted skill only for its originating family raised the mean hidden trajectory utility from $0.713$ (global memory) to $0.816$, and eliminated harmful deployments (from six out of eight to none).
Across 27 paired randomized‑order streams, Scoped‑ORC improved mean trajectory utility by $0.063$ (95% CI $[0.037, 0.094]$) over Global‑ORC, accepted 63 updates instead of 12, produced multiple accepted updates in 19 of 27 streams, and recorded 0 harmful acceptances out of 63. The global control’s utility of $0.713$ fell below the static agent’s $0.775$ because locally valid edits interfered with unrelated families.
These findings establish scope matching as a complementary control for persistent agent memory: certification determines whether an edit is supported, while retrieval scope determines where that evidence authorizes its use.
Review