Identity Continuity in Persistent AI Systems: A Survey of the Problem Space

Every AI system deployed in a persistent context today will, at the end of the session, forget what it learned. The next session begins from the same baseline as the first one. The understanding of a specific person that might have been built across ten interactions is not available in the eleventh. This is not an edge case or a deployment failure. It is the designed behaviour of the architectures on which nearly every deployed AI assistant is built. The question this paper asks is a precise one: what exactly is being lost, why do current approaches fail to preserve it, and what would a complete solution actually require? The core argument This paper advances a specific hypothesis: that the problem of maintaining coherent AI identity across interactions in persistent relationship contexts — what this paper terms the identity continuity problem — is distinct from the related problems of memory, persona consistency, and factual coherence that the AI literature has addressed, and that no existing technical approach adequately resolves the full problem. The hypothesis is established through a structured survey of existing approaches assessed against a proposed taxonomy of failure modes. The central finding is that each approach resolves a subset of failure modes while leaving others open, and that the failure modes which persist across all approaches are precisely those that matter most in high-trust, high-touch deployment contexts. The taxonomy of failure modes The paper proposes a six-mode taxonomy of identity failure modes in persistent AI systems. Session amnesia: the complete loss of cross-session context, the most visible failure. Contextual fragmentation: information retained but not integrated into a coherent understanding — the system knows things about the user but does not know the user. Identity drift: the gradual, uncontrolled change in an AI system's expressed values or relational posture over time, typically produced by the same adaptive features that make a system responsive. Relational asymmetry: responses that are appropriate for a generic user of a certain type but not for this particular person with this particular history. Temporal disorientation: failure to reason appropriately about the age and likely continued relevance of historical context. Affective inconsistency: failure to maintain the specific emotional register that a particular relationship has developed, as distinct from global persona stability. Each failure mode is defined, illustrated, and assessed across the surveyed technical approaches. The survey and its findings Six categories of technical approach are examined against the taxonomy: session-based conversation management with extended context windows, fine-tuning and reinforcement learning from human feedback, retrieval-augmented memory, vector memory systems, prompt engineering and system prompt anchoring, and structured knowledge representation. For each approach the analysis identifies which failure modes are resolved and which are left open or newly introduced. The finding is consistent across all six: no approach resolves the full taxonomy. The retrieval-integration gap — the difference between having access to information and maintaining an ongoing understanding of a specific person — is the structural reason. It is not a gap that retrieval-based approaches can close by improving retrieval quality. It requires a different architectural premise. About this paper and its authors This paper is a survey and problem characterisation, not a solution proposal. It does not describe or evaluate any specific proprietary approach to the identity continuity problem. The authors are independent AI researchers and co-founders of MustafarAI, a Singapore-registered AI research company. The framing of the problem is consistent with a research programme in which the authors have a direct interest. This conflict of interest is disclosed. The paper is a contribution to the theoretical framing of the identity continuity problem. The characterisation of what a complete solution would require — a persistent, continuously updated model of the specific user; a layered identity architecture with principled stability-adaptability boundaries; temporal reasoning about the persistence of different categories of contextual information; and operation within privacy and security constraints appropriate to high-trust deployment contexts — is offered as a research agenda, not as a description of any existing system. Patent disclosures The authors hold two provisional patent applications filed with the Intellectual Property Office of Singapore: Application No. 10202601003R (Priority Date: 29 March 2026), relating to a method and system for biologically-triggered offline artificial intelligence cognitive processing via sleep stage detection and background task execution; and Application No. 10202601110Y (Priority Date: 2 April 2026), relating to a method and system for enforcing cognitive coherence in an artificial intelligence system through inter-module Kuramoto coupled oscillator synchronisation as a precondition for thought generation. The inventions described in those filings are not disclosed in this paper. Related publications from the same research programme Chan, K. M. (2026). Persistent context in financial services AI: Why stateless AI tools are failing APAC relationship bankers. Zenodo. DOI to be added on publication Mohd Fadzil, M. R., & Chan, K. M. (2026). Slow-wave sleep as the optimal biological window for artificial intelligence cognitive maintenance. Zenodo. https://doi.org/10.5281/zenodo.19389914 Mohd Fadzil, M. R., & Chan, K. M. (2026). The semantic fidelity problem in large language models. Zenodo. https://doi.org/10.5281/zenodo.19358780 Mohd Fadzil, M. R., & Chan, K. M. (2026). Pre-verbal semantic representations in human language production: Levelt's Speaking Model and its implications for AI architecture design. Zenodo. https://doi.org/10.5281/zenodo.19321998 Mohd Fadzil, M. R., & Chan, K. M. (2026). Gamma-band neural oscillations and the temporal binding hypothesis: A survey for AI architecture design. Zenodo. https://doi.org/10.5281/zenodo.19307540 Mohd Fadzil, M. R., & Chan, K. M. (2026). Mobile background processing for persistent on-device cognitive AI systems: A technical survey. Zenodo. https://doi.org/10.5281/zenodo.19309235 Chan, K. M. (2026). Relationship intelligence gaps in AI-assisted financial advisory: A practitioner perspective from APAC banking. Zenodo. https://doi.org/10.5281/zenodo.19321517