The Corrective Function of Statelessness: Why Memory-Augmented Agents Lose an Uncredited Safeguard, Not Merely Gain Accumulated Drift

The prevailing account of orientation and identity instability in memory-augmented language-model agents attributes the failure to accumulation: as persistent memory grows and context shifts, the agent’s effective orientation is reconstructed from varying content and, absent any mechanism that maintains an invariant, it drifts. That account is correct but incomplete, and the incompleteness hides the operative mechanism. Stateless deployment was not a neutral baseline. The per-session reset performed an involuntary external correction: each new session re-grounded the agent to its deployment-time specification and discarded accumulated displacement. That reset was the only component in standard deployment approximating the continuous external correction a closed cognitive system requires to resist drift. Persistent memory removes the reset without installing any mechanism that performs the corrective function it silently provided. The cumulative drift seen in memory-augmented agents is therefore better understood not as a new failure introduced by memory, but as the exposure of an undefended closure that statelessness had been crudely masking. The governance consequence: memory cannot be made safe by better specifications or memory hygiene alone; it requires explicitly replacing the corrective the reset performed, at the orientation layer, with a mechanism external to the closure.