Systems Regenerate Faster Than They Remember

Memory benchmarks test whether LLMs can recall facts when directly asked. But in real conversation, users don't quiz the AI—they expect it to naturally weave prior context into responses without prompting.

Method: A 4-month deployment with 40 users and 1,872 sessions tested 7 memory conditions. Direct QA accuracy ranged from 19.7% to 70.1%, but user satisfaction didn't budge. The disconnect: a system scoring 78.8% on Direct QA referenced only 7.9% of those facts in actual conversation—a 71-point gap. Natural integration (detecting relevance and weaving context unprompted) correlated with satisfaction; Direct QA didn't.

Caveats: Tested on general conversation. Task-specific domains (customer support, medical) may show different integration patterns.

Reflections: What architectural changes enable models to detect relevance and integrate context naturally, rather than just retrieve on demand? · Does natural integration degrade over longer time horizons (6+ months) as memory volume grows? · Can users learn to trust natural integration, or do they revert to explicitly prompting for recall?

Developers write security rules in CLAUDE.md using natural language ("do not access X"), but have no way to know if those rules are enforced by built-in controls or merely interpreted by the model.

Method: Analysis of 481 public CLAUDE.md files found that only 4.4% (95% CI: 2.6-6.7%) of extracted security rules had a matching built-in control under strict matching criteria. The range was 4-16% depending on matching strictness. Two security practitioners independently verified a sample with close agreement. This is a write-only channel: developers write rules but get zero feedback on enforcement mechanism.

Caveats: Extraction method captured 66.3% of eligible rules. Actual coverage of all security rules may differ.

Reflections: Can tooling automatically flag CLAUDE.md rules that lack built-in enforcement and suggest control mappings? · How often do model-interpreted rules fail in practice compared to built-in controls? · Should agent platforms surface enforcement status in the configuration UI to close the feedback loop?

Generative UIs regenerate components on demand, but ignore the structural logic users have already established—how items are grouped, ordered, and labeled. Without persistent structure, each generation drifts from what users built.

Method: Maru captures user prompts and interactions as Information Architecture preferences across four elements: partition (grouping), hierarchy (nesting), order (sequence), and vocabulary (labels). These persist as rules that both user and system reference across generations. A user study showed IA persistence kept generated UIs aligned as sessions progressed, while alignment without persistence degraded. Diverse patterns emerged across users and contexts.

Caveats: Study focused on information tasks. Applicability to other GenUI domains (dashboards, forms) unverified.

Reflections: How should systems resolve conflicts when new user requests contradict established IA rules? · Can IA preferences transfer across different information domains or remain context-specific? · What's the right granularity for IA rules—per-session, per-project, or global user preferences?