Subtract, Transport, or Replay? Auditable Deletion from Language-Model Memory

Limited Memory Language Models (LMLMs) externalize factual knowledge to a database to enable deletion-based unlearning without retraining. Existing evaluations measure post-deletion correctness in aggregate and cannot tell whether a deleted fact persists through residual parametric memory, alternative retrieval paths, or near-neighbor retrieval artifacts. We propose a causal auditing framework that holds the model fixed and varies the database state at inference time across three interventions: FULL, DEL-ON, and DEL-OFF. The framework decomposes post-deletion behavior into parametric leakage L(f), retrieval-mediated correctness R(f), and a retrieval artifact rate grounded in the inference-time retrieval trace. We apply it to 12,228 alias-closure deletions across thirteen databases, including four adversarial topologies (Base, Alias, Noise, Collision) we construct in three domains, and six prompt formulations. Parametric leakage is near zero in every variant and every prompt style: the model rarely returns the deleted answer in the absence of retrieval. The residual that does survive lives in the retrieval graph: retrieval-mediated correctness and the retrieval artifact rate match within rounding everywhere, so post-deletion correctness is, in our audit, predominantly reconstituted from near-neighbor retrieval. This residual ranges from 0.7% on the released LMLM database to 13.6% on the most adversarial variant, and prompt formulation does not independently control how much of a deleted fact survives. These results suggest that, for this class of LMLM and deletion procedure, the unlearning boundary is drawn primarily by the database administrator rather than by the model.

Delta-rule linear attention improves recurrent memory updates by correcting what is already stored at the current write address before writing new content. However, the active correction is still anchored to that same write address. As a result, stale information stored at a different address cannot be actively removed before new content is written elsewhere. We propose Erase-then-Delta Attention (EDA), a memory update rule that decouples where to erase from where to write. The key insight is that recurrent memory models should not only correct the current write, but also selectively suppress outdated memory at an independently chosen address. Concretely, our method first applies a targeted erase step along a learned erase direction, and then performs the standard delta-style corrective write along the current write direction. This preserves the corrective behavior of delta-rule updates while expanding their memory-management capacity. Language-model pretraining experiments across dense 2.5B and MoE 25B-A2.8B model families show that EDA performs best in both settings. The gain persists after 80B-token long-context midtraining of the MoE models, where EDA also performs best in long-context evaluations from 4k to 128k contexts. A compact update analysis and memory-state probes suggest why: EDA keeps the delta-rule corrective write intact while allocating an additional cleanup path most strongly when passive decay is weak. These results suggest that recurrent memory models should decide not only what to write, but also what stale information to erase and where.

A language model's memory can be worse than no memory at all. Give a model a memory that kept a wrong conclusion but dropped the work behind it, and it re-emits the stale value as a confident answer; give the same model an empty memory, and it abstains. We call this failure brittle memory. The information loss behind it is definitional (an answer cannot be recomputed once its inputs are gone), so the loss is only the setup; the finding is behavioral. Whether a model turns the lost source into a confident error or an abstention is set by disposition, not capability: four of eight models we test emit, and the four that abstain escape only by an interface affordance -- forced through a mandatory structured-output field, as production tool calls are, they commit the inherited wrong value. We measure correctability with reclaim evaluation: induce a known drift, compress the interaction at a fixed budget, deliver a correction that names the error, and score exact recovery of the known answer, judge-free. Correctability is bottlenecked not by capability but by whether the memory kept a re-derivation basis (the source) rather than the answer, so an 8B model and a frontier one wall in the same place. A one-line source-first policy -- keep the recomputable source, drop the re-derivable conclusion -- restores correctability at equal budget wherever the source is compact and identifiable, with a length-matched control that rules out 'more text' and a deployable one-prompt form weaker than the oracle. We map where the fix fails (source size, noise, a silent truncation mode a completeness tag makes loud), show the failure compounds through memory loops, and replicate on three deployed memory systems and on real dialogue (MultiWOZ, where the checkable value is present by construction). We release the harness, the paired memory conditions, and validators built to come out false.