Do agents need to sleep?

Abstract: LLM-based agents, which increasingly operate persistently across sessions and accumulate memory through interaction, already exhibit the familiar symptoms of unchecked plasticity: semantic drift, degraded retrieval precision, propagation of stored errors, and saturation of memory capacity. Drawing on the selectionist "ringing, resonance and renormalization" (RRR) account of sleep, we propose that each of its three phases maps onto an operation feasible in digital memory architectures. Engineers building stateful agents, meanwhile, are independently rediscovering the problems sleep evolved to solve: what to forget, and when to do the forgetting. From this we derive the falsifiable prediction that selectionist consolidation will outperform instructionist alternatives as memory stores grow large. The core claim concerns the nature of continuous learning itself: plasticity has a price, and for any system that learns continuously under finite capacity while coupled to its environment, that price is sleep.

• Agent memory as a stability–plasticity problem

Central thesis

“Do agents need to sleep?” argues that persistent AI agents face a structural problem analogous to the biological stability–plasticity dilemma. Agents that continuously encode interactions into finite-capacity memory systems eventually accumulate redundant, contradictory, and erroneous traces. The resulting degradation is not primarily a failure of the underlying LLM, but a consequence of unregulated plasticity in the external memory substrate. The paper’s central claim is therefore architectural: any system that learns continuously under finite capacity while remaining coupled to its environment requires sleep-like periods of partial disconnection for memory consolidation and renormalization (2611.02216).

The argument is deliberately stronger than the observation that agents may benefit from periodic context compression. Context-window limitations can be addressed by summarization or retrieval engineering, but the paper maintains that these techniques do not eliminate the underlying cost of continual encoding. New memories can interfere with consolidated knowledge, irrelevant records can reduce retrieval precision, and iterative transformations can cause stored representations to drift away from their original referents. Consequently, memory maintenance must be treated as a recurrent consolidation process rather than as an ancillary retrieval operation.

Persistent-agent memory as a stability–plasticity problem

The paper identifies several failure modes that emerge when agents accumulate experience across sessions. An agent may exhibit “experience-following” behavior, in which its outputs increasingly track retrieved records in proportion to their similarity to the current input. This mechanism can propagate errors stored in memory into later decisions. Semantic drift arises when memories are repeatedly summarized or updated, with each transformation potentially increasing divergence from the original information. As irrelevant entries accumulate, retrieval precision declines; as incompatible records coexist, contradictions become harder to detect and resolve.

These effects imply that an agent’s effective competence can decline even when its parametric model remains unchanged. The memory system alters the signal-to-noise ratio of the information available to the model and can thereby reduce performance through retrieval-mediated contamination. The paper appropriately frames this as a general consequence of continuous encoding rather than as an implementation-specific defect. However, the generality of the claim depends on the memory architecture: systems with immutable provenance, strong contradiction handling, or restricted write policies may exhibit different degradation profiles from naïve append-only stores.

The paper connects these engineering problems to synaptic homeostasis. Wakefulness is characterized, in the biological account, by net potentiation driven by environmental interaction. Potentiation supports learning but incurs energetic and representational costs, increases interference, and eventually approaches capacity limits. The analogous digital process is the accumulation and strengthening of memory traces during online operation. In both cases, plasticity increases adaptability while simultaneously making the system more vulnerable to interference and saturation.

This framing establishes an important implication: increasing memory capacity or expanding the context window may postpone degradation without solving the consolidation problem. If the system continues to encode noisy, redundant, and mutually inconsistent traces, larger storage can merely increase the volume of material requiring selection and raise the complexity of retrieval.

The RRR model as a computational account of consolidation

The proposed mechanism is derived from the selectionist account of sleep known as ringing, resonance, and renormalization (RRR). During ringing, neural circuits are comprehensively sampled through spontaneous activity rather than selected only by an active task or external query. Resonance evaluates which circuits produce coherent, recurrent activity. Renormalization then weakens less coherent synaptic configurations while preserving connections embedded in strongly resonant circuits.

The paper maps these phases onto operations in digital memory systems:

• Ringing: stochastic, query-independent traversal of stored memory traces, including novel combinations that are not directly prompted by current interaction.
• Resonance: evaluation of semantic or structural coherence among co-occurring traces, potentially using vector similarity, graph consistency, or recurrent retrieval patterns.
• Renormalization: selective pruning, deduplication, contradiction resolution, abstraction, and compression of episodic records into higher-order representations.

The principal contribution of this mapping is not any individual operation. Similar operations already exist in agent-memory systems. Rather, the paper emphasizes that they should form a unified cycle in which broad, relatively unbiased memory sampling generates candidates for coherence evaluation, and coherence then determines which traces are retained or compressed. Existing systems often implement only local, instruction-driven updates, such as comparing a new fact with the most similar stored entry and selecting ADD, UPDATE, DELETE, or NOOP. Such procedures are computationally tractable because digital memories are discrete and addressable, but they do not necessarily provide global competition among stored traces.

The distinction between selectionist and instructionist consolidation is consequential. An instructionist system applies an externally specified decision to one new item at a time. An RRR-like system allows stored traces to compete based on their participation in coherent configurations. The paper predicts that selectionist consolidation will outperform instructionist alternatives as memory stores become large and highly interconnected (2611.02216). The proposed rationale is that the combinatorial structure of global memory consistency will eventually exceed the capacity of any explicit local scoring rule, whether hand-designed or learned.

This is a clear and falsifiable claim, but its evaluation requires careful operationalization. “Performance” must include not only downstream task accuracy but also memory growth, retrieval precision, contradiction rates, provenance preservation, update latency, and computational cost. A selectionist mechanism that improves consistency by aggressively deleting rare but valuable information would not necessarily be superior. The paper’s hypothesis therefore requires benchmarks that distinguish useful abstraction from destructive forgetting.

Why disconnection is part of the mechanism

A central point is that consolidation cannot be fully equivalent to online memory maintenance. During interaction, the agent must adapt to current inputs and outputs, which biases memory selection toward immediate relevance. Comprehensive down-selection requires a period in which the memory store is evaluated independently of the current environmental demand. If incoming queries continue to alter the store during consolidation, the competitive assessment of traces can be contaminated by ongoing writes and changing retrieval priorities.

The paper accordingly treats partial environmental disconnection as functionally necessary rather than as a biological accident. “Sleep” in the proposed digital analogue need not reproduce mammalian slow waves, spindles, sharp-wave ripples, or REM-specific events. It denotes a computational regime in which interaction-driven updates are suspended or isolated while memory is sampled, evaluated, and renormalized.

The proposed architecture permits graded forms of disconnection. A lightweight online process could maintain responsiveness, while an offline process operates on a snapshot or isolated partition of memory. This resembles systems that divide memory maintenance into immediate updates and deferred reconciliation, including duplicate merging and contradiction resolution. Local sleep-like processes could consolidate individual partitions while the agent continues serving queries elsewhere.

The qualification is important: partition-level maintenance does not establish that the entire memory system has undergone comprehensive consolidation. If cross-partition dependencies are central to coherence, isolated processing may miss contradictions or redundancies distributed across partitions. The paper therefore distinguishes practical local sleep from full consolidation, which would require at least temporary isolation of the relevant memory substrate from environmental interaction.

Relation to existing agent-memory strategies

The paper situates its proposal within a broader movement toward persistent LLM agents. Existing memory systems address storage, retrieval, summarization, updating, and long-term personalization, but they differ in whether maintenance is local or global, online or offline, and instructionist or emergent. “Memory in the Age of AI Agents” surveys this design space and documents the increasing importance of memory management for stateful agents (Hu et al., 15 Dec 2025).

The paper’s RRR framework provides a conceptual criterion for comparing these systems. A method should be assessed not merely by whether it retains relevant facts, but by whether it periodically exposes the memory store to broad sampling, evaluates coherence across traces, and selectively reduces representational load. Under this criterion, ordinary retrieval augmentation is insufficient: retrieval selects information for a query, whereas consolidation selects information for continued storage.

The comparison with Mem0 is especially direct. Mem0 performs fact-level comparison against similar memories and uses an explicit ADD, UPDATE, DELETE, or NOOP policy (Chhikara et al., 28 Apr 2025). This offers practical advantages, including interpretability, controllable write behavior, and efficient local updates. The paper’s objection is not that such methods are ineffective at modest scale, but that local decisions may fail to capture higher-order dependencies once the store becomes large and densely interconnected. The unresolved empirical question is whether global or stochastic competition delivers enough benefit to offset its computational and systems complexity.

Empirical status and strength of the claims

The paper is primarily a theoretical and programmatic argument. It synthesizes biological sleep theory with observed failure modes in persistent AI-agent memory and proposes a testable architectural hypothesis. It does not report a new controlled experiment, benchmark score, or ablation demonstrating that RRR-based consolidation already outperforms instructionist memory management.

Accordingly, the paper’s strongest numerical implication is not an achieved performance result but a scaling prediction: the advantage of selectionist consolidation should increase as memory stores grow larger and more interconnected. This prediction should be tested under controlled memory growth, with matched storage budgets and comparable update frequencies. Relevant outcome measures would include retrieval precision, task success, contradiction persistence, semantic drift, error propagation, memory footprint, and consolidation overhead.

The biological evidence invoked by the paper supports the plausibility of sleep-dependent synaptic down-selection across diverse organisms, including mammals, fish, insects, worms, and jellyfish. It also supports the existence of localized and asymmetric sleep-like states. Nevertheless, biological universality does not by itself establish computational necessity for digital agents. The paper’s claim becomes testable only after “sleep-like state,” “comprehensive sampling,” “resonance,” and “renormalization” are defined as measurable algorithmic properties.

Limitations and open questions

The paper leaves several technical questions unresolved. First, it does not specify a complete RRR algorithm. Stochastic traversal, semantic resonance, and selective compression are plausible primitives, but the paper does not define their sampling distributions, stopping criteria, coherence objective, or update rule. Different implementations could produce substantially different behavior.

Second, semantic similarity is not equivalent to epistemic validity. Cosine similarity can identify related traces while failing to distinguish corroborated information from repeated misinformation. A robust resonance mechanism may require provenance, temporal structure, source reliability, causal consistency, and task-conditioned utility in addition to embedding alignment.

Third, finite capacity is central to the thesis, but digital systems can often expand storage elastically. The relevant constraint may therefore be effective retrieval capacity, latency, energy, or model context rather than literal storage exhaustion. The claim that sleep is necessary should be tested under explicit resource constraints rather than assumed from biological analogy.

Fourth, the paper does not resolve how consolidation should handle low-frequency but high-value memories. Selectionist competition could favor frequently retrieved or highly interconnected traces and eliminate information that is rarely accessed but critical under unusual conditions. Any practical system would need retention safeguards, provenance preservation, and mechanisms for recovering from erroneous pruning.

Finally, the paper does not establish whether complete disconnection is required for global consistency in realistic agent architectures. Snapshot-based or partitioned consolidation may be sufficient if memory dependencies are sparse, but insufficient if memories form a densely coupled knowledge graph. This remains an empirical systems question rather than a conclusion supplied by the biological analogy.

Conclusion

“Do agents need to sleep?” presents continuous learning as a resource-management problem governed by the cost of plasticity. Persistent agents accumulate memories that can improve adaptation but also generate interference, semantic drift, retrieval noise, and capacity saturation. The paper proposes RRR as a computational framework in which stochastic memory sampling, coherence-based selection, and selective renormalization are integrated into periodic consolidation, ideally under partial environmental disconnection.

Its principal testable claim is that selectionist consolidation will eventually outperform local instructionist memory updates as stores become large and interconnected. The paper does not yet provide the algorithm or empirical validation needed to establish that claim, but it identifies a precise research program: evaluate whether sleep-like consolidation improves long-term agent performance under finite resources without simply replacing useful memories with more aggressively compressed abstractions.