Which Self Does AI Remember? LLM Memory and the Plural Self

Current LLM memory systems represent users by extracting observations and consolidating them into a consistent, predictable account of the individual. This memory paradigm implicitly encodes an ontology of the user as a singular, canonical self. This is operationalized through contradiction resolution schemes; when a user says something that contradicts a stored fact, these systems treat it as an error to resolve through given policies (e.g., recency and frequency). We explore contradiction handling in LLM memory systems under a notion of the self as plural, or containing a multiplicity of coexisting and sometimes conflicting activated modes. Through an analysis of contradiction in user memory we suggest that not all conflicting facts are errors -- some are evidence of plurality. In mode alternation, or instances in which users step into a different activated mode of self, resolving conflicts is not error correction, but a system-imposed selection of which self persists. We argue that failing to preserve such plurality can shape which versions of the self remain salient to users. We describe this as the memory-identity loop, where user representations are surfaced back to the user and condition subsequent developments of the user's self-concept that are in turn fed back into the system.