
What this covers
This is an 18-minute conference talk from a MongoDB speaker, delivered to AI engineers, on how to design memory for agentic systems. It starts from the premise that today's applications are stateless and keep nothing between executions. It then borrows from neuroscience, including the Hubel and Wiesel visual-cortex work that later inspired convolutional networks, to argue that agent builders should copy the brain's memory structure. The talk defines an agent as an entity with perception, an LLM for cognition, and tool use, with memory as the most important distinguishing feature. It also treats "agenticity" as a spectrum, like self-driving autonomy levels, and calls the debate over what counts as an agent unproductive.
The middle of the talk is a taxonomy. Beyond short-term and long-term memory, the speaker separates working memory, semantic and episodic memory, and procedural memory, which is tied to the cerebellum. These map onto concrete agent components. Conversation memory carries timestamps, conversation IDs, and recency and association signals. Workflow memory records execution steps, failures included, so later runs can avoid dead ends. Persona memory gives an agent a consistent personality that supports relationships with users. Toolbox memory stores tool JSON schemas in the database and retrieves only the relevant ones, which gets around the limit OpenAI recommends on how many tool schemas to put in context.
The talk then moves to practice. Memory management is described as six components: generation, storage, retrieval, integration, updating, and deletion. Deletion is recast as forgetting, and retrieval is named the most critical piece. The speaker claims that persistent memory can reduce reliance on repeated prompt engineering. The commercial argument is that no single memory approach exists, citing MeGPT, ME Zero, and Zep, so builders need a flexible store. The speaker offers MongoDB for that role, with text, vector, graph, and geospatial retrieval for RAG in one database. The talk also mentions the Voyage AI embedding acquisition and the open-source Memoriz library.
AI agents require sophisticated memory systems modeled on human cognitive architecture to become reliable, believable, and capable; MongoDB provides the database infrastructure to implement these diverse memory types at scale.
- Human intelligence fundamentally depends on memory (short-term, long-term, semantic, episodic, procedural), so AGI must replicate this architecture
- Current stateless LLM applications fail to accumulate information across executions; memory persistence transforms them into stateful agents
- MongoDB's flexible document model and multi-type retrieval (vector, text, graph, geospatial) enables practical implementation of memory management systems
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The ultimate goal of AI development is to create computational entities that either surpass human intelligence in all domains (AGI) or at least mimic human intelligence, making it essential that AI systems replicate the memory mechanisms that determine human intelligence.
“The whole point of AI is to build some form of computational entity that surpasses human intelligence or mimics it. Then AGI, we're focused on making that intelligence surpass humans in all tasks we can think of. And if you think about the most intelligent humans, you know, what determines the intelligence is their ability to recall. It's their memory. So if we if AI or AGI is meant to mimic human intelligence is a no-brainer, no pun intended, that we need memory within the agents that we're building today.”
Nature has created the most effective form of intelligence available today (the human brain), and AI architects should look inward to understand biological memory and intelligence systems when building agentic systems.
“There is a better architect of intelligence. It's nature. Nature's created our brains. It's the most effective form of intelligence and well some humans that I meet. But it's the most effective form of intelligence that we have today. and we could look inwards to build this agentic system.”
MongoDB acquired Voyage AI earlier in the year to acquire the best embedding models in the market, including text, multimodal, and re-ranker models, enabling MongoDB to reduce AI hallucination in RAG and agentic systems.
“we really understand the importance of managing data and managing memory and that's why earlier this year we acquired Voyage AI. Now they create the best no offense open AI embedding models in the market today. Voyage AI embedded models are we have a text multimodel we have re-rankers and this allows you to really solve the problem or at least reduce AI hall elucination within your rag and aentic systems”
Current applications are stateless—they do not maintain state between executions—and this is a fundamental limitation that memory systems must address by enabling state persistence.
“we're going to be talking about the stateless applications that we're bu building today and how we can make them stateful”
The human brain contains multiple distinct memory systems—short-term, long-term, working memory, semantic, episodic, and procedural memory—which collectively determine human intelligence, and procedural memory (skills and routines) is stored in the cerebellum.
“you have different forms of memory, and that's what makes you intelligent. That's what makes you retain some of the information... There is short-term, long-term, working memory, semantic, episodic, procedural memory. Um, in your brain right now, there is something called a cerebellum... that's where you store most of the routines and skills you can do.”
Hubel and Wiesel won a Nobel Prize in the late 1990s for research on the visual cortex of cats, discovering that visual processing works by learning hierarchies of representation (edges, contours, abstract shapes), which later inspired the architecture of convolutional neural networks used in face detection and object detection.
“You know who these two guys are? Damn. Okay, this is Hob and Wiso. They won a Nobel Prize um in the late 90s, but they did some research on the visual cortex of cats... they found out that the visual cortex of the brains between cats and humans actually worked by learning different hierarchies of representation. So edges, contours and abstract shapes. Now people that are in deep learning would know that this is how convol convolutional neural network works.”
MongoDB provides all the retrieval mechanisms needed for RAG pipelines beyond vector search alone, including text search, graph search, and geospatial search, all within a single database.
“you would store all the information of your JSON schema now you'll begin to understand that MongoDB gives you that flexible data model the document data model is very flexible it can adapt to wherever data wherever model you want your data to take wherever structure and you have all of the retrieval capabilities graph vector text geospatial query in one database.”
The debate about what constitutes an agent is unproductive because agenticity is best understood as a spectrum rather than a categorical distinction, similar to the autonomous vs non-autonomous spectrum in self-driving vehicles.
“I don't like to go into that debate because that's like asking what is consciousness um is a spectrum”
There are multiple distinct forms of memory beyond the basic short-term/long-term dichotomy: conversational entity memory, knowledge data stores, cache, and working memory, each serving different functions in agent systems.
“when we talk about memory, the easy way you can think about it is short-term, long-term, but there al other distinct forms, right? Um, conversational entity memory, knowledge, data, store, cache, working memory.”
Workflow memory captures the execution steps of agent tasks, including failures, and stores them as learning experiences that can be retrieved in subsequent executions to inform the LLM to avoid previous failure paths or explore alternative approaches.
“Workflow memory is very important. You build your agentic system, they execute a certain step. Step one, step two, step three, it fails. But one thing you could do is the failure is experience. It's learning experience. You can store that in your database... you can then pull that in in the next execution to inform the LLM to not take this step or explore other paths.”
The AI development community should mirror the approach of bringing neuroscientists and application developers together to solve problems and advance toward AGI, as demonstrated by MongoDB's recent interdisciplinary collaboration.
“we are having these conversation and once again we're mirroring how we're bringing neuroscientists and application developers together to solve and push us on the path of AGI.”
The concept of 'agenticity' (the degree to which a system exhibits agent-like properties) exists on a spectrum, from minimal agents like an LLM running a loop to level-4 autonomous agents with full access to tools and minimal prompting constraints, analogous to self-driving car autonomy levels.
“the agenticity and that's a word now agenticity of uh of an agent is is a spectrum so they're different levels... we have a minimal agent whereas an LLM running the loop great then you have a level four is autonomous agent a bunch of agents that have access to tools you they can do whatever they want”
Memory is critical to AI agents because it enables them to be reflective, interactive, proactive, reactive, and autonomous—and most or all of these properties can be solved through proper memory implementation.
“Memory is important. It's important because we're trying to make our agents reflective, interactive, proactive, and reactive and autonomous. And every most of this, if not all, can be solved with memory.”
Toolbox memory stores the JSON schema definitions of an agent's available tools in MongoDB, allowing the agent to scale beyond OpenAI's guidance of limiting context window to 10-21 tool schemas by retrieving only relevant tools via database queries before invoking the LLM.
“the guidance from OpenAI is you should only put uh the schema of maybe 10 to 21 tools in the context window. But when you use your database as a toolbox where you're storing the JSON schema of your tools in MongoDB, you can scale because just before you hit the LLM, you can just get the relevant tool using any form of search.”
Memory management is the core technical practice that AI memory engineers must focus on, defined as a systematic process of organizing all information being placed into the context window by selectively pulling in relevant memories and structuring them effectively to enable relevant responses.
“the core topic that we are going to be working on as AI memory engineers is on memory management. We're going to be building memory management systems. And memory management is a systematic process of organizing all the information that you're putting into the context window. Yes, we have like large context window, but that's not for you to stuff all your data in. That's for you to pull in the relevant memory and structure them in a way that is effective”
Memory management systems consist of six core components: generation, storage, retrieval, integration, updating, and deletion, though deletion should be reconceptualized as implementing forgetting mechanisms rather than permanent erasure, similar to how humans selectively forget non-traumatic memories.
“the core components of memory management. Generation, storage, retrieval, integration, updating, deletion. There's a lie here because you you don't delete memories. Humans don't delete their memories except traumatic one and you want to forget. But we really should be looking at implementing forgetting mechanisms within the memory management systems that we're building.”
Current prompt engineering practices are over-reliant on repeated prompt construction, and focusing on memory persistence can reduce the need for extensive prompt engineering.
“we're going to be talking about the prompt engineering that we're doing today and how we can reduce that by focusing on persistence”
Persona memory is a distinct memory type that allows AI systems to have personality, making them more believable and human-like, which helps create relationships with users, and can be modeled in MongoDB databases.
“they gave chat GBT a bit of personality right um and they didn't do a good job but they are going in the right direction which is we are trying to make our systems more believable right we're trying to make them more human we're trying to make them create relationship with the consumer with the users of our systems persona memory helps with that and you can model that in MongoD db”
The Memoriz open-source library encapsulates memory implementation and design patterns including multiple memory types, serving as both an educational tool and a practical resource for implementing diverse memory mechanisms in agent systems.
“I'm working on an open source library. I'm ashamed of the name. I was trying to be cool when I came up with it. It's called Memoriz. Um you can type that on Google. You'll find it. But it has all the design patterns of all of this memory that I'm showing you.”
Retrieval is the most critical component of memory management systems, and MongoDB provides this retrieval capability through multiple search mechanisms (not just vector search) that are essential to RAG pipelines.
“the most important bit is retrieval. And I'm getting to the MongoDB part... MongoDB is that one database that is core to rag pipelines because it gives you all the retrieval mechanisms. Rag is not just vector. Vector search is not all you need. You need other type of search and we have that with MongoDB.”
An AI agent is fundamentally defined as a computational entity with three core properties: awareness of its environment through perception, cognitive abilities through an LLM, and the ability to take action through tool use, with memory (short-term or long-term) being the most critical distinguishing feature.
“What is an AI agent? It's a computational entity with awareness of his environment through perception, cognitive abilities through an LLM and also can take action through tool use. But the most important bit is there is some form of memory short-term or long-term.”
AI agents can be designed to build relationships with customers by using memory systems that enable the agent to recall and reference previous interactions and maintain a consistent persona across conversations.
“we're going to be turning the responses in our AI application and making our agents build relationship with our customers and all of it is going to be centered around memory”
Conversation memory stores back-and-forth conversational exchanges with LLMs (like ChatGPT or Claude) in databases with metadata including timestamps, conversation IDs, and memory signals (recency, association) to implement selective forgetting mechanisms.
“Conversation memory is a bit obvious, right? Back and forth conversation with uh chat GPT with Claude. You can store that in your database as well in MongoDB as conversational memory. And this is what that would look like. Time stamp, time stamp, and you have a conversation ID and you can see something there called recall recency and associate conversation ID. And that's my attempt at implementing some memory signals.”
MongoDB will integrate Voyage AI's embedding models and re-rankers directly into MongoDB Atlas within a few months, eliminating the need for developers to write custom chunking strategies for their data.
“in a few months we're going to be pulling in Voyage AI, the embedded models and the rerankers into MongoDB Atlas and you will not have to be writing chunking strategies for your um for your data.”
The speaker and Tenu (MongoDB's Chief AI Scientist and Voyage AI founder) recently sat down with neuroscientists including Kenneth (who has been researching human brain and memory for over 20 years) and Charles Parker (creator of MeGPT) to have conversations bridging neuroscience and application development toward AGI.
“Last week Saturday myself and Tenu is the chief AI scientist at MongoDB also the founder of Voyage AI. We sat with this three guys in the middle are neuroscientists. Kenneth has been exploring human brain and memory for over 20 years and and over here is Charles Parker. He's the creator of MEGPT your letter and we are having these conversation”
Tools like MeGPT, ME Zero, and Zep exist in the memory management space, but there is no single universal approach to memory management, so developers need a flexible memory provider (like MongoDB) to build custom memory management solutions.
“There are tools out there that focus on memory management. Um, MEGPT, ME Zero, Zep, they're great tools, but after speaking to some of you folks and some of our partners and customers here, there is not there is there is not one way to solve memory. and you need a memory provider to build your custom solution”
Episodic memory and agent registry systems store the persistent state and capabilities of agents themselves (tools, persona, configuration), enabling multi-agent systems to access and coordinate agent information across the system.
“We have episodic memory, we have long-term memory, we have an agent registry, you can store the information of your agent as well... you can see the agent has tools, persona, all the good stuff.”
Agentic RAG extends basic RAG by giving the agent itself the retrieval capability as a tool, allowing the agent to choose when and how to invoke information retrieval rather than having retrieval executed statically by the system.
“you level up. You go into the world of agentic rag, right? You give the retrieval capability to the agent as a tool. And now we can choose when to call on information.”
Agent memory is defined as the mechanisms implemented to ensure that states persist in AI applications, enabling agents to accumulate information, transform data into memory, and inform subsequent execution steps with the goal of making agents more reliable, believable, and capable.
“Agent memory is the mechanisms that we are implementing to actually make sure that states persist in our AI application. Our agents are able to accumulate information, turn data into memory and have it inform the next ex execution step. But the goal is to make them more reliable, believable, and capable.”
Procedural memory and skill routines (like doing a backflip) are stored with approximately 90% confidence in the cerebellum, and this knowledge structure can be replicated in AI agents.
“I heard it's 90% confidence by the way that is actually it is right. I'm not going to do one but but it's stored in that part of your brain.”
The evolution of AI applications over the past 2-3 years progressed from LLM-powered chatbots (2022: ChatGPT launch), to retrieval-augmented generation (RAG) for domain-specific knowledge, to scaling compute and data which enabled emerging capabilities like reasoning and tool use, and finally to AI agents and agentic systems.
“We started off with chat bots. LMN power chatbots. They were great. Chat GPT came out November 2022 and yeah exploded. Then we went into rag. We gave this chat bots more domain specific relevant knowledge and it gave us more personalized responses. Then we begin to scale the compute, the data we're giving to the LLMs and they gave us emerging capabilities, right? Reasoning uh tool use. Now we're in the world of AI agents and agentic systems”
MongoDB positions itself as the memory provider for agentic systems because it provides developers and AI engineers all the features needed to transform data into memory and make agents believable, capable, and reliable.
“MongoDB is the memory provider for Aentic systems. And when you understand that we provide the developer, the AI memory engineer, the AI engineer all the features that they need to turn data into memory to make the agents believable, capable, and reliable.”
MongoDB's mission is to make developers more productive by removing considerations and concerns around managing different data types and handling chunking and retrieval strategies, by pulling these processes into the database itself.
“the mission for MongoDB is to make the developer more productive by taking away the considerations and all the concerns around managing different data and all the process of chunking in retrieval strategies. We we pull that into the database. We are redefining the database.”
Entity memory is a distinct memory type in agent systems (distinct from the other types described), though the speaker does not elaborate on its specific function or implementation details.
“There's entity memory as well.”
Over the next 6 months, the information provided will be very relevant and will put developers in the best position to build believable, capable, and reliable AI applications and agents.
“this information I'll give you within the next 6 months will be very relevant and it will put you in the best position to build the best AI applications to build the best agents that are believable, capable, and reliable”
MongoDB is positioned as being built for the technological changes that will occur immediately, in the near future, and over the next several years, suggesting its architecture is designed for long-term relevance.
“MongoDB is built for the change that we are going to experience now, tomorrow, in the next couple years.”