When an LLM-based agent decides that a fact in memory is stale and deletes it, it is making a bet. The research shows that the bet can turn into a hidden cost: when the resolver errs and drops a query-critical fact, downstream accuracy collapses. Across three models, a controlled sensitivity intervention moves hard-drop from gains between +0.186 and +0.230 under perfect resolution to losses between -0.176 and -0.181 at the lowest tested reliability. The effect reverses sign: the same rule that helps when the system correctly identifies stale facts becomes harmful when identification is uncertain.

Paired targeted interventions show that false-drop causes significantly greater degradation than false-keep. Removing a fact that was needed costs more than keeping a useless one. But even a non-deleting baseline can fail when incorrect labels steer generation. The issue is therefore not simply deleting less as a matter of principle, but understanding where the risk sits: in the quality of the labels that decide what to keep.

On a frozen 102/102 calibration/test split, a deployment-matched gate reaches 0.655 accuracy, against 0.641 for always hard-drop and 0.662 for a test-informed oracle. The gap between the simple gate and the oracle highlights the cost of not having gold labels: it is not an operational threshold, but a model- and task-specific diagnostic. Global-accuracy crossovers require gold labels and should not be used as deployment rules.

For teams running agents on controlled infrastructure, the message shifts priorities. Memory management is not only a model or context problem; it is a cost decision. The paper suggests treating deletion as cost-sensitive and validating query-critical false-drop risk on data that reflects real traffic. If such validation is unavailable, the default should be label-free retention.

This has second-order implications for self-hosted deployments. In an on-premise environment, where the organization owns the entire pipeline, introducing a deployment-matched gate is easier to control than relying on an opaque cloud service. But it also increases the validation burden: representative calibration data and a measure of resolver reliability are needed, otherwise the conservative default becomes a way to avoid errors without understanding the system. It is not free: keeping everything can bloat context and increase inference costs and latency, a trade-off that matters for TCO evaluations. For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks on /llm-onpremise to assess these trade-offs.

There is also a structural signal. Emphasis shifts from quantization or hardware toward the quality of metadata and memory policies. An agent that wrongly deletes facts is a system that cannot trust its own memory; data sovereignty and local control become levers for building custom validation without external labels. But the paper also reminds us that a non-deleting system can fail if labels are wrong: the real vulnerability is supervision quality, not just filter aggressiveness. The next generation of memory systems for agents will have to decide not whether to forget, but what evidence is needed to do so without breaking queries. Until that evidence exists, not deleting is the cheaper move.