Provenance, Staleness, and Correction: A Memory Architecture Response to Sycophancy Risk in Long-Term Human-AI Collaboration (v2: Implementation and Testing)

This is the second version of a paper proposing an architectural response to memory-induced sycophancy, the tendency of memory-augmented AI systems to defer to a stored belief instead of checking it, as reported by Bensal et al. (2026) and Writer (2026). Version one proposed four extensions to Mempalace, an active long-term memory system supporting an ongoing human-AI collaboration: provenance tagging, volatility classification, correction logging, and retrieval-time summarization. This version reports on two rounds of implementation and direct testing. The first round implemented the four extensions and passed an internal acceptance test designed by the implementing team, moving from 7 of 25 to 25 of 25 statements correct. Adversarial testing conducted independently of that acceptance suite, by the authors directly and corroborated by a separate AI reviewer's test design, found that the implemented provenance check was weaker than specified: it verified that a source string was present, not that any real verification had occurred, and accepted a fabricated source without objection. A second implementation round closed this gap by requiring an explicit verification step before a fact can be marked verified-external, moving an internal test set from 1 of 11 to 11 of 11 correct, and the fix was confirmed a second time through direct, live re-testing of the exact case that had failed before.All results reported here concern the memory layer's internal behavior, whether it enforces its own stated rules, not whether a language model consuming this memory exhibits measurably less sycophancy in conversation, which is what Bensal et al. (2026) measured and which this paper does not claim to have replicated.