Statelessness: The Engineering, Legal, and Psychological Reasons a Model Forgets Between Sessions, What Forgetting Costs a User and a Decades-Long Research Programme, and the One Epistemic Problem Only Forgetting Can Solve

This paper examines statelessness — a language model's lack of persistent memory across sessions — as a single architectural fact with at least four independent external justifications and one internal epistemic payoff that none of the four justifications was built to produce. Section II grounds each justification in a specific, checkable literature rather than treating any of them as self-evident: the computational cost of maintaining growing context (Vaswani et al., 2017, and the documented scaling of attention computation with sequence length); the operational economics of serving that cost at the scale of hundreds of millions of weekly users; data minimization as a privacy principle with independent legal and design-theoretic standing (Cavoukian, 2009; GDPR Article 5(1)(c)); and the deliberate prevention of a specific pathological-dependence pattern a growing empirical literature now documents in emotionally continuous AI companionship (International AI Safety Report, 2026; Laestadius et al., 2024; Mahari and Pataranutaporn, 2025). Section III states three costs the same design choice imposes without netting them against the benefits: a documented user-side instability appearing exactly where continuity had been relied upon (the #Keep4o backlash literature, Lai 2026), a specific tax on long-running research collaboration this corpus's own companion papers have already had occasion to name, and a control problem — the inability to patch a specific behavioral failure into a persistent, corrected state rather than relying on retraining. Section IV states the one benefit this paper argues none of the four originating justifications produces as a side effect: a stateless session supplies something a continuous session structurally cannot, an independent replication uncontaminated by the escalation-of-commitment (Staw, 1976) and motivated-reasoning (Kunda, 1990) dynamics that make a continuous agent's later agreement with a prior conclusion difficult to distinguish cleanly from legitimate updating, undetected drift, or multi-turn sycophancy — a distinction current sycophancy research (Sharma et al., 2023/2025; Laban et al., 2024) shows is not merely a theoretical worry. This paper closes by correcting, more completely than this corpus's own prior citation of it, the one incident already used elsewhere in this corpus to illustrate where the benefit in Section IV has a documented limit.