According to a paper published on arxiv.org, researchers have introduced memory canonicalization, a framework designed to address how different Large Language Models (LLMs) interpret identical stored memory objects differently.
The research addresses a growing challenge in persistent LLM memory systems. According to the paper, systems such as MemGPT/Letta, Mem0, and Zep now provide agents with “tiered, temporally-aware, model-agnostic external storage,” while the Model Context Protocol (MCP) standardizes access to memory servers. However, the researchers note that “an identical stored memory object, retrieved by two different LLMs under otherwise identical conditions, may not be interpreted the same way, factually or emotionally.”
The proposed solution is a “write-time pipeline that detects ambiguity, conditional structure, and emotional loading in a raw memory object and rewrites it into an explicit, structurally disambiguated canonical form, with emotional valence represented as a separate field rather than inferred from tone,” according to arxiv.org.
The researchers tested their framework using 176 synthetic memory objects across three model families, developing a Cross-Model Semantic Drift / Emotional Consistency Score benchmark (CMSC-E). According to the paper, they found “an uncorrected improvement in cross-model emotional consistency for fully canonicalized memory relative to raw memory (+0.050, 95% bootstrap CI [0.013, 0.086], paired t-test p = 0.010).” However, the researchers acknowledge this result “does not survive Bonferroni, Holm, or Benjamini-Hochberg correction” and that none of the factual-drift comparisons reached significance. The authors characterize their results as “exploratory rather than confirmatory.”