Draft:Agent memory

In artificial intelligence, agent memory refers to the mechanisms used by autonomous agents to store, retrieve, and manage information across interactions, tasks, or sessions. Large language models (LLMs) are generally stateless, meaning that information from previous interactions is not retained unless it is supplied through an external memory system. In AI agents, memory is used to store, retrieve, and manage information beyond the limits of a model's context window, enabling continuity across tasks and sessions.

Agent memory systems are commonly divided into working, episodic, semantic, and procedural forms, although classifications vary among architectures and research frameworks. These systems may be implemented using retrieval mechanisms such as vector databases, knowledge graphs, and other external storage architectures. Agent memory is used in applications including personal assistants, software engineering agents, and research agents, and remains an active area of research due to challenges involving scalability, retrieval accuracy, privacy, and memory management.

Definition and scope

"Memory plays a pivotal role in enabling large language model (LLM)-based agents to engage in complex and long-term interactions," according to Zeng et al. Agent memory encompasses the storage, retrieval, and management mechanisms that allow artificial intelligence agents to maintain information beyond a single interaction or context window. In LLM-based agents, memory is typically implemented through external storage systems rather than within the language model itself. These mechanisms support the preservation of state, context, and task-relevant information beyond individual steps.

The scope of agent memory in this article is limited to memory systems used in autonomous or semi-autonomous AI agents, particularly those based on large language models. It focuses on external or architectural memory mechanisms used during inference and does not cover model training processes or internal parameter-based learning.

Background

The development of agent memory emerged in response to limitations of large language models in long-term and complex interactions. Since the emergence of transformer-based large language models in 2017, these systems have operated within fixed context windows, which limits access to prior interactions unless reintroduced in the prompt.

While language models can process information provided within a context window, they do not inherently retain information across interactions. This limitation became more prominent with the widespread deployment of large language models in 2022, when conversational and task-oriented applications began requiring continuity across multiple sessions. As autonomous agents increasingly undertook long-horizon tasks requiring the accumulation and retrieval of information over time, researchers developed external memory systems to support persistence, adaptation, and continuity of behavior. Reinforcement learning agents typically do not use persistent, queryable memory across episodes, instead relying on state representations or experience replay mechanisms developed in the 2010s.

Cognitive architectures in AI agents have also drawn on earlier concepts from cognitive science such as Soar (developed in the 1980s) and ACT-R (first proposed in 1993, with major revisions into the early 2000s) and modeled human cognition using structured representations of memory and reasoning, including distinctions among working, episodic, semantic, and procedural memory.

The Cognitive Architectures for Language Agents (CoALA) framework, introduced in 2023, provided an early formalization of memory as a modular component within language-agent architectures, contributing to the framing of memory as a distinct subsystem in LLM-based agents.

Subsequent research in the mid-2020s has increasingly treated agent memory as a design layer in autonomous systems rather than a property of the underlying language model.

Types of agentic memory

Agent memory systems are commonly categorized into several functional types based on their role in storing, organizing, and retrieving information within artificial intelligence agents. These categories are not strictly standardized and may vary across implementations and research frameworks.

Working memory

Working memory refers to information actively maintained during an agent’s current reasoning or task execution. In large language model-based agents, this is typically represented by the model’s context window and any short-term intermediate reasoning state used during inference.

Episodic memory

Episodic memory stores records of specific past interactions, events, or experiences encountered by an agent. This may include prior conversations, task trajectories, or sequences of actions and outcomes, often indexed with temporal or contextual metadata to support retrieval.

Semantic memory

Semantic memory refers to generalized knowledge extracted from past interactions or external sources, representing facts, concepts, or stable information that is not tied to a specific event. In agent systems, this information is often consolidated from episodic memory or external knowledge bases.

Procedural memory

Procedural memory represents learned behaviors, strategies, or action sequences that guide how an agent performs tasks. In AI agents, this may include tool-use policies, reasoning patterns, or reusable workflows that can be invoked across tasks.

These categories are conceptually influenced by distinctions in cognitive science, but in artificial intelligence research they are primarily used as functional design abstractions rather than strict psychological analogues.

Architectural implementations

Agent memory in artificial intelligence systems is implemented using a variety of architectural approaches that typically combine external storage systems with retrieval and integration mechanisms within large language model-based agents.

Retrieval-based memory systems

A common architectural approach is retrieval-based memory, in which information from prior interactions or external sources is stored in an external database and retrieved based on relevance to the current input. Retrieval is often performed using similarity search over vector representations of stored information, enabling approximate matching between queries and past records. These systems are typically implemented using external storage backends, such as dedicated vector databases, or multi-model document-oriented databases like MongoDB that natively support both vector search and structured application data.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation (RAG) is frequently used as a foundation for agent memory systems. In these architectures, external information is retrieved and incorporated into the model’s context window prior to generation, allowing the model to condition its output on stored or retrieved knowledge. In agent settings, RAG may be extended to include conversational history, task-specific state, or structured records of prior actions.

Structured memory systems

Some agent memory systems use structured representations such as knowledge graphs or relational databases to encode entities, events, and relationships between stored items. These approaches enable explicit relational querying and support more structured reasoning over stored information compared to unstructured retrieval methods.

Hybrid memory systems

Hybrid memory architecture (HMA) combines multiple memory representations, typically integrating unstructured vector-based retrieval with structured storage mechanisms. As Zhao et al. observe, "existing approaches face a fundamental trade-off between efficiency and effectiveness: memory compression risks losing critical details required for complex reasoning, while retaining raw text introduces unnecessary computational overhead for simple queries." This design is used to balance flexible semantic retrieval with explicit relational structure, particularly in systems requiring multi-step reasoning or long-horizon task execution.

Memory consolidation and management

Many agent memory systems include mechanisms for consolidating or compressing stored information over time, such as summarization, filtering, or pruning. These processes reduce storage overhead and improve retrieval relevance by retaining salient information while discarding redundant or low-utility entries over time.

Integration into agent reasoning loops

Retrieved memory is typically reintroduced into the input context of a language model during inference. This allows external memory to influence generation, reasoning, and action selection within a given task without modifying the model’s internal parameters.

Challenges

Agent memory systems introduce a range of technical and conceptual challenges related to retrieval accuracy, information management, system scalability, and privacy in autonomous artificial intelligence agents.

Retrieval and memory quality

The effectiveness of agent memory depends on the accurate retrieval of relevant information. Memory systems may return semantically related but contextually inappropriate information, while the accumulation of large volumes of stored data can reduce retrieval precision. In addition, agents may retain outdated or contradictory information, creating inconsistencies between retrieved memories and current circumstances.

Memory management and scalability

Long-term memory systems must balance the retention of useful information against storage and computational constraints. Techniques such as summarization, consolidation, and pruning can improve efficiency, but may also result in the loss of detail or the introduction of inaccuracies. As memory stores grow, retrieval latency and infrastructure requirements may also increase.

Privacy and governance

Persistent memory systems raise concerns regarding the collection, storage, and retention of user-specific information. Questions surrounding consent, data ownership, deletion, and access control have become increasingly important as memory capabilities are incorporated into deployed conversational agents.

Standardization and evaluation

Agent memory remains an emerging area of research. As Du observes, "the field still lacks a community-standard evaluation harness," with each benchmark using its own datasets, metrics, and protocols, making cross-paper comparison unreliable. Differences in memory representations, retrieval methods, and benchmarking practices make direct comparison between systems difficult.

• Intelligent agent