Establish the memory-specific contribution of Reconsider

Establish whether the memory-specific components of Reconsider—the five checks, minimal-sufficient-set procedure, four handling modes, and associated demonstrations—produce improvements beyond general conversational, reliability, and role-play guidance.

Background

The primary comparison contrasts the complete Reconsider prompt with a short baseline, so any advantage may result from general guidance, examples, or stylistic instructions rather than from the memory-use framework itself. To probe this issue, the paper constructs General-Matched by removing only Reconsider's memory-specific content while preserving its general guidance and shared scenario instructions.

The matched-control experiment uses only DeepSeek and Claude and yields directionally positive but small and variable margins. Because of the limited and nonrepresentative sample, the observed result does not establish the independent contribution of the memory-specific material; larger experiments and human evaluation are required.

References

Net win rates for Full over General-Matched were directionally positive in every judge $\times$ generator cell, although the margins were small and varied across cells. This preliminary result is consistent with a modest contribution of the memory-specific content beyond the shared general guidance, but given the limited sample sizes it does not establish that contribution. Future work will therefore conduct human evaluation on the same pairs and on larger sets of experiments.

— Before Bringing It Up: When and How AI Companions Should Use Memor  (2610.09470 - Guo et al., 7 Oct 2026) in Section 3.5, Preliminary Matched-Guidance Control