When LLM Agents Should Forget: Quantifying False-Drop Risk in Memory Filtering
Summary:This preprint examines a failure mode in LLM-agent memory: a stale-fact resolver can incorrectly hard-delete a fact that the current query needs. Across three models, hard deletion changes from gains of +0.186 to +0.230 under perfect resolution to losses of -0.176 to -0.181 at the lowest tested reliability. Paired interventions show that false drops degrade downstream accuracy significantly more than false keeps, although retaining facts is not risk-free if incorrect labels can steer generation.
Core Finding
This work does not introduce a new model; it audits the forgetting decision inside an LLM-agent memory pipeline. Directly hard-deleting facts labeled as stale does not automatically improve answer quality. If the staleness resolver makes an error and the deleted fact is critical evidence for the current query, the memory system removes noise at the cost of making the required evidence unavailable.
The central conclusion is not that memory should never be deleted. Rather, the value of deletion depends on the structure of resolver errors. Because deletion occurs before retrieval, a false drop does not merely lower a fact's rank; it removes the fact from the candidate set entirely. Systems should therefore estimate the cost of falsely dropping query-critical facts separately from the cost of falsely retaining stale facts, instead of optimizing average freshness or global accuracy alone.
Mechanism: Why Hard Filtering Backfires
A typical memory pipeline can be expressed as: write facts into memory M → assign stale or current labels with a staleness resolver → hard-filter stale items → retrieve only from M_keep = {m ∈ M | resolver(m) ≠ stale} → generate an answer.
The crucial structural difference is that deletion precedes retrieval. If a query-critical fact m* is incorrectly labeled as stale, it does not merely rank lower; it disappears from the candidate set, leaving the generator no opportunity to use it. Compared with ordinary retrieval noise, this is an availability failure that the downstream model may be unable to repair.
A first-order approximation of the expected harm is:
E[harm] ≈ P(query-critical fact is dropped) × C_FD + P(stale fact is kept) × C_FK
Here, C_FD is the cost of a false drop and C_FK is the cost of a false keep. Paired targeted interventions in the paper show that false drops cause significantly greater degradation than false keeps in the tested settings, indicating that the two error types are not symmetric. This does not establish a universal ordering for every model and task; production systems still need to estimate C_FD and C_FK from their own data.
Query criticality is also local. A fact can have different value across different queries, so high global stale-label precision does not make deletion safe. A small number of false drops concentrated on critical facts can offset the benefit of many correct deletions. The paper also observes that retaining everything is not risk-free: if incorrect stale labels are supplied to the generator and influence its reasoning, the answer may still be wrong. Retaining content and propagating its labels should therefore be treated as separate decisions.
Experimental Results: Deletion Gains Reverse with Resolver Reliability
| Experimental condition | Strategy | Result | Engineering implication | |---|---|---:|---| | Perfect resolution | Always hard-drop | Relative gain of +0.186 to +0.230 | When resolution is reliable, removing stale facts can improve downstream performance | | Lowest tested reliability | Always hard-drop | Relative loss of -0.176 to -0.181 | At sufficient resolver error, hard filtering becomes a major source of harm | | Frozen calibration/test split | Always hard-drop | 0.641 accuracy | Unconditional deletion ignores the unequal risk of individual facts | | Same split | Deployment-matched gate | 0.655 accuracy | Selective deletion improves accuracy by 1.4 percentage points over always dropping | | Same split | Test-informed oracle selection | 0.662 accuracy | The gate approaches a diagnostic reference, but the oracle is not directly deployable |
The absolute gap between the gate and always dropping is 0.014, while its gap from the test-informed oracle is 0.007. These results support selective forgetting over one-size-fits-all deletion, but the stability of the 1.4-point gain still requires validation on larger and independent samples.
Evaluation Boundaries and Objectivity
The available material provides only the arXiv v1 abstract, so the evidence remains at the preprint stage. First, the split contains 102 calibration examples and 102 test examples. The abstract reports no confidence intervals, repeated-run variance, or detailed significance tests, making the 0.014 gap potentially sensitive to sample composition.
Second, the abstract does not identify the three evaluated models, task types, memory-write policy, resolver implementation, label-construction process, or exact decision rule behind the deployment-matched gate. Models may differ in their sensitivity to incorrect labels, and deployment-matched data should not be interpreted as a universal threshold.
Third, the reported global-accuracy crossover depends on gold labels and varies by model and task, making it a diagnostic result rather than an operational switch. A production system cannot obtain gold labels for every live query, so an aggregate accuracy crossover cannot be copied directly into deployment.
Fourth, the test-informed oracle selects a policy using test information and should be treated only as a reference baseline. It may overestimate the attainable ceiling and can carry optimistic bias from test-set information, so it is not a deployable solution.
Engineering Trade-offs
| Strategy | Main benefit | Main risk | Deployment requirement | |---|---|---|---| | Always hard-drop | Compact context, lower Token use, cleaner retrieval candidates | Resolver errors irreversibly remove critical evidence | Extremely high and stable resolver reliability | | Always retain | Avoids content false drops and keeps rollback simple | Stale facts pollute context, incorrect labels may steer generation, and Token use and latency increase | Sufficient context budget and prompt isolation | | Deployment-matched gate | Highest deployable accuracy in the reported experiment | Requires matched labels, calibration maintenance, and drift monitoring | Reproducible calibration data and task-level evaluation | | Retain and downrank | Preserves recoverability and reduces hard-deletion exposure | Incorrect facts may still be retrieved, ranking complexity increases, and the paper does not directly evaluate it | A stable soft-ranking signal and independent validation |
The final strategy is an engineering alternative rather than a method proven by this paper. Even when soft downranking is used, uncalibrated stale labels should not be injected into the generation context as strong instructions, because that can reproduce the label-steering failure observed in the study.
Developer and Deployment Guidance
-
Instrument the fact-query path first. Log memory snapshots, resolver labels, deletion events, query content, final task outcomes, and tool-call results. Without these fields, teams cannot distinguish a false drop from an underlying model failure.
-
Annotate query-critical false drops separately. Each evaluation record should indicate whether a fact is stale, whether it is critical to the current query, and how task success changes after deletion. If annotation is too expensive, teams can estimate criticality through controlled ablation: temporarily remove one fact and observe whether the answer or tool behavior degrades. This approach is more expensive and noisier than direct labeling.
-
Calibrate the gate only on deployment-matched data. The calibration set should cover realistic memory sizes, query distributions, resolver versions, and context budgets. Freeze the test set after choosing the threshold; never tune the policy backward using test outcomes.
-
Report accuracy alongside operational costs. Track task success or accuracy, query-critical false-drop rate, stale-fact retention rate, retrieval Tokens, end-to-end latency, context occupancy, and deletion recoverability. A global average can conceal low-frequency but high-cost failures.
-
Default to label-free retention without matched validation. This is the paper's conservative recommendation. A system can first retain content without propagating stale labels, then reduce noise through ranking or context isolation rather than physically deleting facts without evidence.
-
Monitor resolver drift continuously. Track false drops and false keeps by fact type, tenant, and time window; use canary releases, automatic rollback, and manual audits. The memory composition and query distribution can change, requiring recalibration rather than reuse of an old threshold.
-
Prefer recoverable deletion where possible. Versioned tombstones or short retention windows can replace immediate physical deletion, allowing critical facts to be restored after a false drop. The added storage and governance costs must be evaluated separately.
— END —Source: ArXiv AI (cs.AI) (2026-11-16)
Tags: #LLM Agents #Long-Term Memory #Memory Filtering #False-Drop Risk #Cost-Sensitive Decision
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments