From Memory Accumulation to Governed Experience: A Longitudinal Case Study and Retrospective Ablation of Effective Capability in Human–LLM Work Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Long-running AI assistants can appear to become more capable even when user-specific model-weight learning has not been established. This longitudinal single-case study examines that phenomenon at the level of the human–LLM work system. Version 2.1 adds and refines retrospective empirical analyses of archived comparative data: a 38-generation context-boundary experiment, a blinded structured-interaction benchmark, closed-checklist versus open-anomaly quality-assurance evidence, and machine-browser/fresh-context verification records. Under a strict post-hoc recoding of the 38-generation archive, hard-violation rates were 2/10 under minimal context, 6/10 after adding supported but incomplete context, 3/8 when absence of personal-history information was explicit, and 0/9 when absence of requested specific facts was explicit. A broader sensitivity coding that includes two borderline unsupported-specificity cases yields 3/10, 7/10, 3/8, and 0/9 respectively. These comparisons are descriptive, not causal, because coding was post-hoc and prompt sets differed. A separate blind benchmark showed that a detailed-spec direct-generation method was preferred over a fixed structured planner in 9 of 14 reviewed comparisons, while the planner retained stronger knowledge-boundary control. The combined evidence rejects a simple “more memory = more capability” account. The best-fitting systems hypothesis is Externally Governed Experience Accumulation (EGEA): selected external memory combined with provenance, positive and negative boundaries, tests, tools, independent verification, and real-world feedback. The evidence is convergent and operationally meaningful but is not a randomized causal demonstration of the full stack. A pre-specified isolated-session replication protocol is included as a supplement.