8 min readJust now

–

Abstract

Memory is what allows intelligence to persist over time. In the human brain, memory enables learning from experience, maintaining a stable sense of identity, and planning actions that extend beyond the present moment. In artificial systems built on large language models, memory plays an increasingly similar role by transforming stateless text generators into agents that can adapt, personalize behavior, and improve through experience.

This paper shows that modern memory mechanisms in LLM-based agents closely parallel biological memory systems. Both rely on short-term memory to support immediate reasoning and long-term memory to preserve knowledge and experience. Both selectively encode information, abstract repeated experiences into general knowledge, retrieve memories in a goal-directed way, and store information in structured, layered forms rather than as raw data. By examining memory formation, evolution, retrieval, and storage across brains and agents, the paper highlights a shared set of principles that underlie persistent intelligence in both natural and artificial systems. These parallels suggest that memory is not merely an auxiliary component of intelligence, but its foundational infrastructure.

2. Memory Formation: Encoding Experience into Representations

2.1 Memory Formation in the Brain

In the brain, memory formation begins with encoding. Sensory inputs are processed in modality-specific cortices and coordinated by attention and executive control networks. The hippocampus plays a central role by binding distributed sensory, spatial, temporal, and emotional signals into coherent episodic representations. These newly formed memories are highly plastic and context-dependent, relying on hippocampal indexing to remain accessible.

At this stage, memory is fragile. Without reinforcement or relevance, representations decay rapidly. Only information that is repeatedly accessed, emotionally salient, or behaviorally relevant is selected for consolidation. This selective encoding ensures that the brain does not store raw experience indiscriminately but transforms it into structured representations optimized for future use.

2.2 Memory Formation in LLM Agents

LLM agents exhibit a functionally analogous, though not mechanistically equivalent, process to biological memory formation. Raw interaction streams, including user inputs, tool calls, observations, and intermediate reasoning states, initially reside within the agent’s working context. This corresponds at a functional level to biological short-term or working memory: both are capacity-limited substrates that support online cognition and must be selectively consolidated to persist beyond the immediate task.

Because the context window is limited, agent architectures implement explicit mechanisms to determine which elements of experience are preserved beyond the current trajectory. In practical agentic AI systems, this selection is realized through concrete algorithms rather than autonomous biological processes. Trajectory-based agents often extract episodic records at task boundaries, storing structured representations of goals, actions, observations, and outcomes. This is analogous to hippocampal binding, in which multimodal inputs are integrated into discrete event representations.

To manage long interaction histories, many systems employ hierarchical summarization, progressively compressing raw experience into more abstract representations. While engineered and task-driven, this process is functionally reminiscent of hippocampal event segmentation, in which continuous experience is partitioned into meaningful episodes. Similarly, importance-weighted memory formation assigns priority based on signals such as novelty, reward, or prediction error. These mechanisms are inspired by, but not equivalent to, the role of neuromodulatory systems in biasing biological memory encoding toward behaviorally relevant events.

Memory formation in LLM agents is therefore not passive logging but an active transformation. Behaviorally relevant experiences are distilled into episodic records describing attempted actions and their outcomes, or into semantic records encoding stable facts, preferences, or rules. This transformation parallels, at a conceptual level, hippocampal indexing processes that convert continuous experience into structured representations optimized for later retrieval. In both cases, memory emerges from selective abstraction rather than exhaustive retention, though in agents this process is explicitly engineered rather than biologically emergent.

3. Memory Evolution: From Episodes to Knowledge

3.1 Memory Evolution in the Brain

Biological memory is dynamic rather than static. Episodic memories evolve over time through consolidation and abstraction. Repeated exposure to similar situations allows the brain to extract regularities, gradually transforming context-bound episodes into decontextualized semantic knowledge. Neurobiologically, this process reflects the gradual transfer of memory dependence from the hippocampus to distributed neocortical networks.

Importantly, retrieval itself reshapes memory. When an episode is recalled, it becomes temporarily labile, allowing it to be updated, strengthened, or weakened before being stored again. Forgetting plays a complementary role, pruning low-value or redundant memories to preserve cognitive efficiency.

3.2 Memory Evolution in LLM Agents

Agent memory systems display the same evolutionary pressures. Early experiences are stored as episodic traces tied to specific tasks or sessions. Over time, agents abstract these traces into reusable strategies, guidelines, or skills. Successful trajectories are generalized into workflows, while repeated factual observations are distilled into semantic memory entries.

In agentic AI, this evolution is realized through reflection and abstraction algorithms. Reflection-based learning mechanisms transform failures into linguistic constraints stored in memory, enabling behavioral adaptation without modifying model weights, much like human learning from mistakes. Experience distillation algorithms summarize multiple trajectories into generalized heuristics or policies, corresponding to the extraction of semantic knowledge from repeated episodes. Skill induction frameworks compile frequently successful action sequences into reusable procedures, closely paralleling the transition from episodic experience to procedural and semantic knowledge in the brain. Forgetting mechanisms such as decay functions or capacity-bounded memory pools prune low-value memories, serving the same interference-reduction role as biological forgetting.

This mirrors biological systems consolidation. Episodic agent memories initially guide reasoning directly, but as abstraction proceeds, they give rise to compact semantic representations that can be applied across tasks. Retrieval-triggered updates further align with biological reconsolidation: when an agent recalls a memory during reasoning, it may revise or reinforce it based on new outcomes. Forgetting mechanisms, such as memory decay or pruning, serve the same role as biological forgetting, preventing uncontrolled memory growth and interference.

4. Short-Term and Long-Term Memory and Their Interaction

4.1 Short-Term and Long-Term Memory in the Brain

Short-term memory in humans supports the temporary maintenance and manipulation of information over seconds. It relies on distributed activity across sensory cortices and frontoparietal control networks, maintained through persistent neural firing or activity-silent synaptic states. Its capacity is limited, forcing prioritization and selective attention.

Long-term memory, in contrast, stores information over minutes to decades. It is supported by hippocampal–neocortical coordination, with episodic memory depending strongly on the hippocampus and semantic memory residing primarily in the neocortex. The two systems interact continuously: short-term memory serves as a workspace for encoding and retrieval, while long-term memory provides priors that shape perception and reasoning.

4.2 Short-Term and Long-Term Memory in LLM Agents

In LLM agents, the context window functions as a form of short-term or working memory. It maintains the immediate conversational or task state and supports real-time reasoning. Like biological working memory, it is capacity-limited and often exhibits recency effects, with more limited and context-dependent primacy effects. Persistence beyond the active context is achieved through long-term memory mechanisms, typically implemented as external memory stores or, in some systems, through updates to model parameters that allow information to persist across sessions and tasks.

In modern agent architectures, the interaction between short-term and long-term memory is governed by explicit control algorithms rather than autonomous biological processes. Techniques such as context folding and dynamic summarization compress the contents of the working context during extended tasks, prioritizing information most relevant to current goals. While engineered and task-driven, these mechanisms are functionally analogous to attentional prioritization in humans, in that both selectively allocate limited cognitive resources.

Memory scheduling or retrieval modules determine which long-term memories are injected back into the context window based on relevance, similarity, and task state. This bidirectional exchange is inspired by, but not equivalent to, hippocampal–prefrontal interactions observed in biological cognition. Retrieved memories shape ongoing reasoning, while the contents of short-term processing influence what is selected for longer-term storage.

In some systems, knowledge that is repeatedly retrieved or reinforced may be internalized into model parameters through continual fine-tuning or reinforcement learning. This process bears a conceptual resemblance to the gradual consolidation of information into long-term cortical storage, though it differs substantially in mechanism, timescale, and flexibility. Overall, agent memory systems exhibit a structured interplay between transient context and persistent storage that parallels biological memory at a functional level, while remaining explicitly engineered and algorithmically constrained.

5. Memory Retrieval: Reinstating the Past for Present Use

5.1 Memory Retrieval in the Brain

Memory retrieval in humans is a reconstructive process. Rather than replaying stored data, the brain reinstates distributed patterns of neural activity associated with a memory. The hippocampus acts as an index, reactivating neocortical representations that together recreate the remembered event or knowledge.

Retrieval is cue-dependent and context-sensitive. External or internal cues bias which memories are accessed, and the retrieved content is shaped by current goals and prior knowledge. Critically, retrieval opens a window for plasticity, allowing memories to be updated before being re-stored.

5.2 Memory Retrieval in LLM Agents

Agent memory retrieval follows the same principles. Queries issued during reasoning act as cues, triggering similarity-based or structured retrieval from memory banks. Retrieved memories are not replayed mechanically but integrated into the current context, influencing reasoning and decision-making.

In agentic systems, retrieval is implemented through similarity-based and structured algorithms. Vector similarity search retrieves memories whose embeddings best match the current query, functioning analogously to hippocampal pattern completion. Graph-based retrieval traverses relational memory structures to reconstruct multi-step experiences, resembling navigation through cognitive maps. Hybrid retrieval strategies combine recency, importance, and similarity signals, closely paralleling how emotional salience and contextual relevance bias recall in humans. Retrieved memories are often re-summarized or updated after use, mirroring biological reconsolidation.

As in the brain, retrieval is selective and goal-driven. Retrieved memories may be revised or re-summarized based on their utility, mirroring reconsolidation. This ensures that agent memory remains adaptive rather than static, continuously aligning stored experience with current objectives.

6. Memory Storage: Locations and Formats

6.1 Memory Storage in the Brain

Biological memory storage is distributed rather than localized. Short-term memory relies on transient activity across sensory and frontoparietal networks, while long-term memory depends on coordinated hippocampal–neocortical systems. Importantly, memories are not stored as raw sensory data but as structured representations, including event-based units and cognitive maps that encode relationships and abstractions.

Synaptic plasticity provides the physical substrate for long-term storage, embedding experience into connection strengths that shape future neural dynamics.

6.2 Memory Storage in LLM Agents

Agent memory storage exhibits direct analogues. Short-term memory resides in the context window, implemented through attention states. Long-term memory is stored externally in memory banks or internally in model parameters. Storage formats range from natural language text and graphs to latent vector representations.

Agent memory storage mechanisms directly reflect these biological principles. Text-based memory stores preserve episodic narratives and align with event-based representations. Graph-structured memories encode entities and relations, functioning similarly to hippocampal cognitive maps. Parametric memory embeds knowledge into model weights through learning, closely mirroring synaptic plasticity and enabling fast, automatic access. Latent memory architectures maintain fixed sets of high-dimensional vectors that are iteratively updated and retrieved, resembling distributed neural codes rather than symbolic storage. Together, these layers form a hierarchical memory system analogous to short-term activation, hippocampal indexing, and neocortical consolidation in the brain.

7. Conclusion

Across memory formation, evolution, retrieval, and storage, the parallels between LLM agents and the human brain are striking. Both systems confront the same fundamental challenge: how to preserve relevant experience over time without overwhelming limited processing capacity. The solutions, layered memory systems, selective consolidation, cue-driven retrieval, and structured abstraction, are convergent.

Understanding agent memory through the lens of cognitive neuroscience does more than provide metaphorical insight. It offers concrete design principles for building more adaptive, efficient, and robust artificial intelligence. As agent architectures continue to evolve, the analogy between brains and LLM agents is likely to deepen, not as a claim of equivalence, but as evidence that memory is the universal infrastructure of intelligence.