Agent-Controlled Forgetting: Selective Memory Management
- Agent-controlled forgetting involves selective and controlled operations on memories such as semantic loss, modification, and deletion, often driven by constraints like finite capacity, privacy, and relevance, and aims to manage and suppress dynamic knowledge accurately rather than indiscriminate erasure.
- Catastrophic, anterograde, and intentional forgetting are key types noted: catastrophic forgetting involves damage to previously learned knowledge from new learning, anterograde forgetting where retaining old knowledge impedes learning new knowledge, and intentional forgetting as selective suppression or deletion of obsolete or unsafe information.
- Several neural learning mechanisms are used for forgetting, including Controlled Forgetting Networks, Cycled Memory Networks, Task Scheduling as a Forgetting Control and Query-conditioned forgetting.
Agent-controlled forgetting is the deliberate regulation of what an artificial agent retains, uses, suppresses, compresses, deletes, or restores. It encompasses several distinct operations: semantic loss of knowledge in possible-world models, selective modification of neural representations, task scheduling, private unlearning, memory pruning, behavioral policy control, execution-state reconstruction, and reversible suppression of obsolete knowledge. The common objective is not indiscriminate erasure, but control over the influence of information under constraints such as finite capacity, privacy, safety, relevance, latency, energy, and changing environments.
1. Conceptual foundations and distinctions
Knowledge forgetting in epistemic models
A foundational formalization treats forgetting as a change in an agent’s Kripke model rather than as contraction of a plausibility-ordered belief set. For an atomic proposition , forgetting whether adds epistemic alternatives in which is true and false. For a complex proposition , there may be several equally minimal ways to alter constituent atoms so that becomes false. The operator
therefore quantifies over all minimal forgetting operations: must hold after every admissible way of forgetting whether . The construction uses prime implicates , falsifying one clause at a time while preserving other atomic facts whenever possible (Fernández-Duque et al., 2015).
The framework distinguishes forgetting whether from the weaker forgetting 0. The former removes knowledge of both 1 and 2, whereas the latter guarantees only that the agent no longer knows 3; it may continue to know 4. These operators are not equivalent, and sequentially applying weaker forgetting does not reproduce the stronger uniform operation.
This formal distinction is relevant to agent control because the primitive forgetting operator is universal rather than choice-based. Individual action-model modalities represent selected clause-falsifying events, while the universal operator represents guarantees that survive every admissible minimal operation. Uniform forgetting applies the same atomic modification across a copied model; dependent forgetting allows the reason for forgetting to vary by world and is therefore more flexible but can produce greater information loss.
Catastrophic, anterograde, and intentional forgetting
In neural continual learning, catastrophic forgetting denotes damage to previously learned knowledge caused by new learning. Anterograde forgetting is complementary: retained or transferred old knowledge obstructs acquisition of new knowledge through capacity shrinkage and conceptual confusion (Peng et al., 2021). A system can therefore preserve earlier tasks while learning later tasks poorly, or learn later tasks while destroying earlier competence.
Agent-controlled forgetting adds a third concern: the system may intentionally suppress or delete knowledge because it is obsolete, unsafe, private, redundant, harmful, or irrelevant. This is distinct from accidental loss. It also differs from simple memory eviction: the decision may depend on semantic relevance, provenance, task utility, user instructions, privacy requirements, or the possibility of later restoration.
Storage, use, and execution state
Several recent frameworks separate different layers of memory:
- Source storage: observations, documents, trajectories, or records retained for possible future use.
- Answer-time or working view: information selected for a particular query or task.
- Persistent cognitive state: a bounded representation that governs future behavior.
- Execution state: summaries, plans, tool traces, KV-cache contents, and other runtime artifacts.
- Parametric state: information encoded in model weights.
This separation is essential. A fact can remain stored but be excluded from the current memory view; a plaintext record can be deleted while its content survives in a summary or cache; and a model can be unlearned parametrically while recovering the target through tools. RD-Forget explicitly preserves a source archive while constructing a query-conditioned memory view 5, so “forgetting” primarily controls use rather than storage (Li et al., 9 Sep 2026). Execution-state unlearning instead requires the runtime to behave as if the target had never been observed, including removal of derived influence from summaries, plans, tool calls, memory writes, and KV-cache state (Yao et al., 4 Sep 2026).
2. Neural and continual-learning mechanisms
Controlled Forgetting Networks
Controlled Forgetting Networks (CFNs) implement selective overwriting in fixed-capacity unsupervised spiking neural networks (Allred et al., 2019). The primary architecture is a single fully connected layer of excitatory leaky integrate-and-fire neurons receiving Poisson-encoded MNIST pixels. Each output neuron represents an approximately local prototype, and lateral inhibition creates competition.
Novelty is detected when ordinary neurons produce insufficient activity during a recognition interval. A dopaminergic neuron then stimulates candidate neurons and temporarily increases their STDP learning rates. Lateral inhibition restricts the effective response to one or a few winners. The selected neurons rapidly move their incoming weight vectors toward the novel input, while other representations remain comparatively stable.
The targeting mechanism favors infrequently used representations. Dopaminergic synapses to frequently active neurons are depressed, reducing their likelihood of future stimulation; underused neurons retain stronger dopaminergic stimulation and are more likely to be overwritten. Thus, the system preferentially repurposes outlier or low-use prototypes. It does not guarantee permanent preservation: once outliers are exhausted, useful generalized representations may also be replaced.
In the sequential MNIST protocol, classes were presented as ten disjoint tasks without replay or task-boundary signals. The 6400-neuron system achieved 6 final test accuracy, and sequential learning incurred an average accuracy reduction of 7 percentage points relative to the corresponding interleaved system. The evidence indicates reduced interference and gradual degradation rather than complete immunity from forgetting.
Cycled Memory Networks
Cycled Memory Networks (CMN) address both catastrophic and anterograde forgetting by separating short-term and long-term memory (Peng et al., 2021). The long-term network, L-Net, stores consolidated historical knowledge; the short-term network, S-Net, learns the current task without having to allocate all capacity to previous tasks. A transfer cell selectively passes features from L-Net to S-Net, while a distillation mechanism transfers selected current-task knowledge back into L-Net.
The transfer cell includes memory processing, a recall gate, and memory integration. The gate depends jointly on historical and current features, allowing irrelevant or conflicting information to be suppressed. After S-Net learns the task, L-Net is updated using both long-memory preservation distillation and short-to-long integration distillation.
CMN therefore separates three controls:
- Plasticity: S-Net provides dedicated capacity for new tasks.
- Transfer: the gate controls whether old features influence current learning.
- Consolidation: distillation determines what enters durable memory.
CMN is not explicit task deletion. Its principal form of forgetting is suppression of interference and selective omission during transfer. The architecture does not provide an operation such as deleting a task-specific memory or erasing a skill from shared parameters.
Task scheduling as forgetting control
In multi-task reinforcement learning, forgetting can be controlled by deciding which tasks to train, when to revisit them, and how much training to allocate to each (Speckmann et al., 3 Mar 2025). Experiments with PPO agents in MiniGrid found human-like forgetting curves: in six of ten seeds, repeated relearning of SimpleCrossing increased the time required for subsequent forgetting, while four seeds exhibited a more periodic pattern.
Leitner-inspired scheduling used five stages, promoting correctly solved tasks and returning incorrectly solved tasks to the first stage. SuperMemo-inspired scheduling used reward quality to determine revisit intervals. Neither consistently outperformed Prioritized Level Replay (PLR), which prioritizes value-prediction error. PLR was the most stable method and achieved the highest final mean performance, while SuperMemo was especially effective on some tasks such as GoToDoor and DoorKey.
The central finding is that task relations are directional. Training Unlock benefited DoorKey more than training DoorKey benefited Unlock; GoToDoor supported GoToObject asymmetrically; and a task’s current performance did not necessarily indicate its future retention. Effective scheduling therefore requires a directed model of transfer and retention rather than a single scalar forgetting score.
3. Memory-centric forgetting in language agents
Bounded cognitive state
The Agent Cognitive Compressor (ACC) replaces transcript replay with a bounded Compressed Cognitive State (CCS) that is rewritten at every turn (Bousetouane, 15 Jan 2026). External artifacts remain available for retrieval, but they are treated as candidate evidence rather than automatically committed memory. A qualification gate filters recalled artifacts before a schema-constrained compressor updates the next CCS.
The CCS schema includes episodic trace, semantic gist, focal entities, relational map, goals, constraints, predictive cues, uncertainty, and references to retrieved artifacts. The design separates:
8
This implements selective non-persistence: information may remain in an external store but be omitted from the persistent state that conditions future reasoning. The approach targets context growth, hallucination carryover, memory poisoning, constraint erosion, and behavioral drift. The supplied evaluation reports bounded memory and qualitative reductions in hallucination and drift, but no concrete implementation of the qualification function, token budget, or systematic ablations.
Adaptive decay and consolidation
FadeMem uses a dual-layer hierarchy consisting of a Long-term Memory Layer and a Short-term Memory Layer (Wei et al., 26 Jan 2026). Each memory has content, textual representation, strength, creation time, and access frequency. Importance combines semantic relevance, access frequency, and recency:
9
Memory strength follows an adaptive exponential decay:
0
with importance-dependent decay rate. Long-term memories decay sublinearly, while short-term memories decay superlinearly. Access reinforces memories with diminishing returns.
FadeMem also uses LLM-guided conflict resolution, distinguishing compatible, contradictory, subsuming, and subsumed memories. Related memories can be fused into consolidated representations, while newer information suppresses older contradictory information. On LTI-Bench, it retained 1 of critical facts while using 2 of storage, corresponding to approximately 3 storage reduction. Its ablations associate performance gains with dual-layer storage, fusion, and conflict resolution.
Structured survival and pruning
SF-AMS models memory survival through Composite Importance Scoring, hierarchical layers, and a survival potential (Yang et al., 29 May 2026). Importance combines LLM-derived semantic salience with entity-level structural priors. Memories are assigned to Core, Important, Secondary, or Irrelevant layers.
The survival update is:
4
Repeatedly retrieved, important, and nonredundant memories gain potential; unused or redundant memories decay. When capacity is reached, the lowest-potential item is evicted. This creates a closed loop:
5
On LongMemEval-s, SF-AMS achieved 6 accuracy, compared with 7 for A-MEM. Removing strategic forgetting reduced performance across all reported task categories, indicating that hierarchical organization alone was insufficient.
Query-conditioned forgetting
RD-Forget treats forgetting as query-dependent evidence selection rather than global deletion (Li et al., 9 Sep 2026). The source archive 8 preserves observations, while the query-conditioned view 9 is selected under a rate-distortion objective:
0
subject to
1
Same-slot replacement links suppress superseded values for current-state questions while preserving them for historical or evolution-oriented queries. This prevents both under-forgetting, in which obsolete values remain active, and over-forgetting, in which historical evidence is destroyed.
In the reported ablations, removing forgetting caused deficits of 2, 3, and 4 percentage points on AMB-Text, LME-KU, and MAB-FC, respectively. Removing query-conditioned curation also caused substantial deficits. These results support the distinction between preserving historical evidence and controlling its influence on a particular answer.
Optical and quantization-based forgetting
ScrapMem organizes multimodal experiences into temporally grounded Scrapbook Pages and applies Optical Forgetting by progressively reducing image quality and resolution with age (Chang et al., 5 May 2026). The Episodic Memory Graph preserves salient entities, events, temporal order, and causal paths while the underlying visual page becomes lower fidelity.
The default Timed-Gentle schedule reduces JPEG quality from 5 for recent pages to 6 for old pages and scales resolution from 7 to 8, with boundaries at 9 and 0 days. In the reported evaluation, storage decreased by up to 1, while forgetting-enabled Joint@10 was 2 and Recall@10 was 3. The mechanism is schedule-controlled optical decay with graph-mediated semantic preservation, not a learned agent policy.
SuperLocalMemory V3.3 uses Ebbinghaus Adaptive Forgetting, lifecycle-aware embedding quantization, consolidation, and seven retrieval channels (Bhardwaj, 6 Apr 2026). Memory retention depends on access, importance, confirmation, emotional salience, and trust. Lifecycle states map to 32-, 8-, 4-, and 2-bit embeddings before deletion. The system distinguishes lossy compression, archival, reduced retrieval probability, and true deletion. It provides automatic lifecycle control and explicit user commands such as slm forget, but does not specify a first-class protocol through which an agent independently chooses whether to protect, compress, or delete an individual memory.
4. Privacy, security, and distributed forgetting
Private unlearning
Continual Learning and Private Unlearning (CLPU) formalizes user-controlled learning followed by exact task deletion (Liu et al., 2022). A request may mark a task as permanent, temporary, or forgotten. Exact unlearning requires the post-deletion model distribution to equal the distribution produced by learning only the surviving request sequence.
CLPU-DER++ achieves this structurally. Permanent tasks are stored in a main model; temporary tasks receive isolated networks and replay memories. Forgetting a temporary task removes its isolated network, replay memory, and status entry without modifying the main model. Since the temporary task never contaminated the permanent model, the remaining state is distributed as if the task had never been learned.
This guarantee is narrow: it depends on temporary knowledge remaining isolated. Once merged into the main model, exact deletion is no longer provided. The method trades storage for exactness by keeping a complete network for each temporary task. On four benchmarks, CLPU-DER++ achieved near-counterfactual privacy according to Jensen–Shannon metrics while generally trailing DER++ slightly in continual-learning utility.
Behavioral and prompt-based unlearning
Secure Forgetting studies LLM-based agents whose underlying LLM parameters are fixed and inaccessible (Ye et al., 1 Apr 2026). The framework distinguishes state unlearning, trajectory unlearning, and environment unlearning. A conversion model translates natural-language requests into operational unlearning prompts, while the target memory entries are physically removed.
State unlearning avoids specified states; trajectory unlearning prevents reconstruction of a sequence while preserving access to individual states; environment unlearning degrades performance in a target environment while preserving unrelated environments. The method is therefore behavioral and memory-level rather than parameter-level.
The reported GridWorld results include state- and trajectory-unlearning efficacy of approximately 4–5, with success rate remaining 6. Environment unlearning increased target-environment steps substantially while leaving other-environment steps near their original values. The method does not establish certified erasure from the underlying LLM and is vulnerable to over-unlearning, under-understanding, and over-reasoning.
Tool-mediated recovery
Agentic Tool Unlearning demonstrates that parameter-level unlearning can fail when an agent can query external tools (Chen et al., 21 Aug 2026). A model may no longer recall a target directly but recover it through web search, databases, retrieval systems, tool observations, or search-query formation.
The framework first performs parametric unlearning and then applies trajectory-level reinforcement learning. Forget-query rewards penalize target-seeking tool use, target-related information appearing in trajectories, and final-answer leakage. Retain-query rewards preserve benign tool use and neighboring knowledge.
For Qwen-3-4B with NPO, standalone forget score was 7, but agent-with-tools score rose to 8; after ATU, it fell to 9. Under forced retrieval directives, however, leakage increased again. This establishes that robust forgetting must regulate the complete information pathway:
0
Distributed co-forgetting
The Co-Forgetting Protocol coordinates memory pruning across multi-agent systems using semantic voting, temporal decay, weighted quorum, and PBFT consensus (Bach, 19 Jun 2025). Agents evaluate memory relevance using DistilBERT and multi-scale decay, then vote to keep or forget. A memory is deleted only if PBFT agrees and the weighted forgetting score reaches the quorum.
The prototype used four agents, two planning agents with weight 1, and two perception/execution agents with weight 2. Reported results include a 3 memory-footprint reduction after 500 epochs, 4 voting accuracy against human annotations, 5 consensus success under one injected Byzantine fault, and an 6 cache hit rate.
PBFT provides agreement and consistency, not semantic correctness, user consent, legal erasure, or reversibility. The prototype remains limited by small scale, synthetic memories, under-specified semantic scoring, store-consistency risks between Pinecone and SQLite, and the absence of a formal appeal or recovery mechanism.
Memory mutation and control-plane placement
Control-Plane Placement Shapes Forgetting distinguishes a recall plane from a control plane (Yang, 14 Jun 2026). The recall plane retrieves facts; the control plane performs supersession, release, and purge. The paper argues that conventional memory benchmarks measure retrieval while missing failures in mutation.
Deterministic systems handled lexical and temporal categories well but failed canonicalization and cross-lingual identity. Inscribe-time LLMs achieved strong canonicalization but could not reliably perform intent-aware deletion. Mutation-time LLM hooks improved deletion of prefix collisions, paraphrases, and compound facts without changing the recall path.
Lethe with a mutation-time LLM achieved 7, and LangGraph with the same type of hook achieved 8, on the in-house adversarial suite. The hook increased mutation latency to approximately 9 seconds per case but left ordinary recall latency unchanged. The architectural conclusion is that semantic intelligence is most useful at mutation time for deletion-sensitive operations, while deterministic execution should validate and commit the resulting bounded mutation plan.
5. Execution-state, reversible, and governance-oriented forgetting
Selective replay of contaminated runtime state
Execution-State Unlearning formalizes forgetting as counterfactual reconstruction of a stateful agent (Yao et al., 4 Sep 2026). Let the runtime transition be
0
If the target first enters at step 1, the prefix through 2 is identical to the counterfactual run. The post-target suffix is conservatively tainted because every later transition may depend on target-contaminated state.
Provenance-Guided Selective Replay locates the target’s source, computes its downstream taint closure, restores a clean checkpoint by cropping the KV cache, sanitizes observations, and regenerates the suffix. Under deterministic decoding, sanitized observations, and replay-safe external effects, the resulting state equals the counterfactual state.
The lower bound for exact replay is
3
state transitions. Selective replay attains this bound, while avoiding the 4 transitions required by a full reset. In the reported experiments, selective replay used 5 recomputed tokens versus 6 for a full reset and was indistinguishable from the full-reset reference in the tested behavioral audits. Plaintext memory deletion left leakage unchanged, and source redaction continued to influence behavior in 7 of episodes.
Reversible forgetting
The Hysteretic Reversible Memory Controller (HRMC) introduces Active, Dormant, and Retired states (Yash et al., 18 Aug 2026):
8
Active memories influence normal behavior. Dormant memories remain recoverable but are suppressed from ordinary retrieval. Retired memories are no longer autonomously restorable, although retirement is distinct from legal or technical erasure.
HRMC smooths relevance evidence,
9
and uses asymmetric thresholds 0, persistence counters, shadow-mode reactivation, and policy-gated retirement. A memory becomes dormant after persistent low relevance or a policy block. It can return to Active only after persistent high relevance and demonstrated shadow-mode benefit. Retirement requires sufficient dormancy and approval.
The framework addresses non-stationary enterprise environments in which a policy, workflow, or market strategy can become harmful under one regime and useful again under another. It does not report completed empirical superiority claims and leaves relevance estimation, shadow gain, authorization, and retirement governance open.
Governance and authorization
A recurring distinction across the literature is between the mechanism that executes forgetting and the policy that decides whether forgetting should occur. Deletion may be requested by a user, administrator, privacy monitor, injection detector, or autonomous component. The execution mechanism can remove or suppress information, but it does not by itself determine authority, scope, legal obligations, or whether derived information should remain.
This distinction is explicit in execution-state replay, where the operator accepts a forget request but does not decide who is authorized to issue it. It is also present in HRMC, where retirement requires policy or owner approval, and in distributed co-forgetting, where PBFT ensures agreement but not semantic validity or user consent.
A governance-capable system therefore requires:
- authorization and scope resolution;
- provenance across source and derived artifacts;
- audit logs recording state transitions and reasons;
- protection against over-deletion and under-deletion;
- reversible quarantine where permanent deletion is not yet required;
- deletion propagation across caches, summaries, indexes, replicas, tools, and other agents;
- verification that the target no longer influences behavior;
- separate handling of legal erasure and operational suppression.
6. Comparative assessment and open problems
Main design paradigms
| Paradigm | Primary control object | Forgetting behavior |
|---|---|---|
| Epistemic forgetting | Accessibility and valuation structure | Adds alternatives representing all minimal knowledge-loss operations |
| Neural controlled forgetting | Local representations and plasticity | Selectively overwrites underused prototypes |
| Continual-learning architecture | Short- and long-term model pathways | Gates transfer and consolidates selected knowledge |
| Memory decay and pruning | Stored memories and metadata | Decays, compresses, demotes, fuses, or deletes memories |
| Private unlearning | Isolated task models and replay state | Deletes temporary knowledge while preserving unrelated state |
| Behavioral unlearning | Prompts, action policies, and external memory | Suppresses target behavior without modifying base weights |
| Execution-state unlearning | Summaries, plans, tools, cache, and runtime state | Reconstructs the counterfactual suffix after a target enters |
| Reversible forgetting | Operational memory status | Suppresses obsolete knowledge while retaining recoverability |
Persistent tensions
Agent-controlled forgetting is governed by several recurring trade-offs.
Retention versus relevance: retaining all historical evidence protects against premature loss but allows obsolete knowledge to influence current decisions. Query-conditioned views, dormant states, and strategic survival scores address this by separating archival retention from operational use.
Privacy versus utility: exact private deletion is easiest when temporary knowledge is isolated before it contaminates shared parameters. Once information is merged into shared representations, deletion becomes approximate or requires expensive retraining and replay.
Safety versus autonomy: automatic systems can suppress harmful or sensitive memories rapidly, but false positives may erase rare safety-critical information. Policy gates, shadow evaluation, auditability, and reversible states limit irreversible decisions.
Semantic flexibility versus determinism: LLMs improve canonicalization, paraphrase handling, and intent interpretation, but introduce variability, latency, cost, and over-deletion risk. Deterministic mutation primitives remain preferable for exact identifiers, TTLs, hard deletion, and audit requirements.
Compression versus fidelity: quantization, optical degradation, summarization, and bounded cognitive states reduce storage and retrieval noise but can destroy details needed for later reasoning. Graph structures, provenance, and query-conditioned extraction mitigate but do not eliminate this risk.
Local control versus distributed consistency: an individual agent can prune its own memory, but shared multi-agent memory requires consensus, synchronized deletion, and protection against Byzantine or inconsistent participants.
Common misconceptions
Forgetting is not the same as refusing to answer. An agent may refuse to emit a target while continuing to act on it, retrieve it, or reveal it under elicitation.
Deleting a source record is not sufficient. Derived summaries, plans, tool traces, cached representations, and parameterized prompts may retain influence.
Reducing retrieval probability is not deletion. Decay, down-ranking, quantization, and dormancy can leave information recoverable through other channels.
Parametric unlearning is not agent unlearning. A tool-augmented agent can reconstruct a supposedly forgotten target from external sources.
PBFT does not guarantee semantic correctness. Consensus ensures that agents agree on a decision, not that the decision is appropriate, authorized, or reversible.
More memory is not always better. Retained knowledge can cause anterograde forgetting, conceptual confusion, negative transfer, stale recommendations, and security exposure.
Agent-controlled does not necessarily mean consciously deliberated by an LLM. In many systems, control is implemented by schedulers, scoring rules, thresholds, replay mechanisms, or policy engines. The agent’s interaction trajectory may influence survival without the LLM explicitly issuing a deletion decision.
Open research directions
Several problems remain unresolved:
- Unified semantics: connecting epistemic forgetting, memory-store pruning, behavioral unlearning, and execution-state counterfactuals.
- Selective provenance: tracking individual attributes and causal dependencies through summaries, compound facts, multi-agent messages, and model-generated plans.
- Certified behavioral non-recovery: moving beyond benchmark probes toward formal or statistically validated guarantees.
- Tool and environment control: preventing recovery through unseen tools, external databases, irreversible actions, or downstream agents.
- Reversible deletion: supporting quarantine, restoration, versioning, and policy-controlled retirement without unbounded dormant storage.
- Adaptive policy learning: learning decay rates, thresholds, transfer relations, and harm models under non-stationary environments without allowing attackers to manipulate the controller.
- Privacy-preserving distributed forgetting: coordinating deletion across agents, vector stores, caches, backups, and cloud services.
- Utility-preserving exactness: identifying when information can be deleted from one task while retaining shared representations needed by unrelated tasks.
- Evaluation: measuring source deletion, derived influence, tool-mediated recovery, behavioral residue, utility preservation, auditability, latency, energy, and cost separately.
- Authorization and governance: determining who may request forgetting, which descendants are in scope, and when an operational retirement must become legal erasure.
Agent-controlled forgetting is therefore best understood as a systems problem. It requires coordination among representation learning, memory architectures, retrieval, task scheduling, policy control, provenance, runtime execution, tool interfaces, privacy mechanisms, and governance. The strongest contemporary approaches do not treat forgetting as a single deletion operation. They regulate the lifecycle and influence of knowledge: retain what remains useful, suppress what is temporarily unsafe or irrelevant, compress what can lose detail, delete what must be erased, and restore what becomes relevant again under controlled evidence.