How does the 'memory' of an AI agent differ from that of a human?

Advances in AI technology are making it possible to create 'AI agents' that can think for themselves, autonomously manipulate and judge external tools and data, and act in accordance with given goals. For AI to perform complex tasks like humans, it needs memory to store the information it is given and retrieve it as needed. Polish engineer brgsk explains how the memory of an AI agent differs from human memory.

Agent memory: an anatomy
https://brgsk.xyz/agent-memory-anatomy/

The memory system of an AI agent is broadly composed of three components: an 'extractor,' a 'storage location,' and a 'search unit.'

The 'extractor' reads the conversation record and extracts only the information deemed useful later as short sentences. This process often uses an LLM (Language Language Module), generating abstract descriptions such as 'Users prefer TypeScript' or 'We prioritize XX in this project.' However, the timing of the extraction is tricky because extracting after every statement would save even casual conversation, while extracting everything at the end of the conversation makes it easy to overlook important information that was in the middle of a longer context.

Furthermore, extraction is the process of compressing 'specific events' into 'facts detached from context.' For example, the event 'I said I liked TypeScript while drinking coffee on Tuesday' can be transformed into the fact 'Users like TypeScript,' but in the process, the date, time, setting, degree of emphasis, and nuance of the statement are lost.

The 'storage location' is a database where the extracted information is stored. Implementation methods include vector indexes for searching by semantic similarity, relational tables for filtering by item, and knowledge graphs for representing the relationships between pieces of information.

The most difficult aspect of data storage isn't deciding where to save the information, but rather how to handle conflicts between old and new information. For example, if a user lived in Paris until April and then moved to Amsterdam, saving both pieces of information as they are would make it appear as if both 'living in Paris' and 'living in Amsterdam' are true. In this case, a design decision is needed: overwrite the old information, leave both and let the search engine decide, or mark the old information as 'currently invalid.'

The 'search engine' is responsible for finding memories that fit the current context when a user asks a question and passing them to the LLM (Lookout Manager). It's fundamentally a vector search, typically combined with keyword search and re-ranking. This structure is very similar to RAG (Search Enhancement Generator) , which uses document search to reinforce answers. The difference is that the search target is not general documents, but rather short descriptions accumulated about users and tasks.

Furthermore, the memories that are stored are classified into categories such as ' episodic memory ,' ' semantic memory ,' ' procedural memory ,' and ' working memory ,' depending on the type of information. Brgsk explains that working memory in AI agents corresponds to the context window and is therefore a different mechanism from the long-term memory discussed here.

'Episodic memory' refers to memories of events associated with a specific time and place. In AI agents, this may be stored as a timestamped description based on conversation records, such as 'The user said they live in Berlin on March 14, 2026.'

'Semantic memory' is the memory of facts that are detached from specific events. Much of what is called the memory of an AI agent is similar to this semantic memory, and is treated as facts such as 'the user prefers TypeScript' or 'the user prioritizes speed in this project.'

'Procedural memory' is memory that is associated with physical actions or behaviors. In humans, this includes actions such as 'riding a bicycle' or 'using keyboard shortcuts,' but this type of memory is not always adequately implemented in the memory libraries of AI agents.

For example, LangMem , a library that enables AI agents to manage long-term memories, has a mechanism to improve system prompts and change behavioral tendencies, making it an implementation close to procedural memory. On the other hand, Mem0 , a memory management system for AI agents, uses the same storage and retrieval mechanisms as semantic memory, and Graphiti, which builds a time-aware knowledge graph for AI agents, does not have the concept of procedural memory at all.

Furthermore, 'prospective memory,' which allows AI agents to recall pre-planned actions at the appropriate time, is also crucial. Brugsk points out that while current implementations have time-specified triggers, they do not adequately realize memory that 'recalls when a condition appears.'

Thus, since episodic memory is compressed into semantic memory during extraction, procedural memory often consists only of labels, and future memory remains largely unexplored, brgsk argues that 'many agent memories are not as broad a memory system as their name suggests.' Extraction is similar to how humans organize memories during sleep, storage is similar to long-term memory, and retrieval is similar to cue-based recall, but actual AI memory is not a biological brain, but rather software that manages states.

The memory function of an AI agent is primarily similar to 'autobiographical memory,' which stores facts about the user. Rather than remembering the agent's own life, it's a system that maintains information on the user's behalf, such as 'where they live, what they work on, who they value, and what decisions they make.'

In recent years, attempts have been made to create 'integration' mechanisms similar to how humans organize memories during sleep. Anthropic's Dreaming and sleep-time compute , proposed by the University of Berkeley and the AI research organization Letta, are technologies that review accumulated memories and past conversations, consolidating duplicates and resolving contradictions to create a more organized storage location.

On the other hand, the human tendency for 'emotionally strong events to be more memorable' cannot be directly transferred to an AI agent that operates solely on text. While there are ways to have LLM score importance, these are merely proxy indicators that estimate emotionality, not emotions themselves.

It's also important to note that memories should not only be retained but also forgotten appropriately. Humans forget because the brain cannot store everything, but for AI agents, the ability to 'correctly find the necessary information' is more important than storage capacity. Therefore, what is required of the memory of an AI agent is not simply to delete old information. The important thing is to design it so that currently valid information is displayed above older information, outdated information is invalidated without being deleted, and past states can be checked when needed.

brgsk pointed out that the memory of an AI agent is not a direct reproduction of human memory, but rather an engineering mechanism for carrying over information obtained from conversations and tasks into future contexts. He argued that the important thing is not to be misled by the word 'memory,' but to understand what it stores, what it loses, and when it can be recalled.