LLM agents rely on memory architectures to store and reuse experience across tasks. Traditional deployments fix these architectures after launch, limiting adaptability when task types change. Recent self‑evolving approaches use execution feedback such as evaluation scores or compressed summaries, but this coarse‑grained feedback cannot pinpoint failures in long‑horizon executions, leading to large, unfocused code rewrites.\ \ MemScope introduces an evidence‑grounded, localized evolution framework. Its Patch Agent jointly inspects raw execution traces, runtime memory state, and the current architecture code to locate repair targets and apply bounded, in‑place edits under provider‑identity and memory‑compatibility constraints.\ \ On the GAIA benchmark, the trace‑and‑memory inspection variant achieves 86.67% accuracy, outperforming trace‑only (70.00%) and summary‑only (63.33%) feedback.\ \ For xBench tasks, the evolved memory architecture raises the correct answer count of DeepSeek V3.2 from 49 to 55 and Kimi K2.7 Code from 50 to 53. On unseen ALFWorld sub‑tasks, success climbs from 50.75% to 71.64%.\ \ Analysis of archived architecture snapshots shows a median code‑change ratio of only 4.4%. These results support fine‑grained execution evidence as a foundation for localized, in‑place memory‑architecture evolution.\ \ Review: MemScope’s combination of trace and state inspection enables precise, minimal edits to memory architectures, markedly improving cross‑task adaptability and highlighting the value of fine‑grained feedback in self‑evolving systems.