Opportunities and Challenges of AI Agent Memory

OpenAI announces the full activation of ChatGPT's memory capabilities ( ChatGPT can provide context based on memory across sessions.

Opportunities brought by memory

The memory function is undoubtedly an important way to increase user engagement with AI applications. Switching AI platforms doesn't usually incur much switching cost, but that changes with memory. For example, the design of Nuwa was developed through conversations with ChatGPT. If I want to discuss anything related to AI Agents, I tend to talk to ChatGPT, as it can understand what the AI Agent on Nuwa can do and how it works without me having to provide too much context.

Therefore, the memory function will be an important direction for all AI applications in the future, including AI Agents. Since everyone is calling on large model interfaces behind the scenes, the gap mainly appears in two aspects:

1. Toolset tools: Can the AI Agent rely on tools to do more things.
2. Memory Ability: Does the AI Agent understand you better?

How to manage memory?

How should the memory of an AI Agent be managed? Taking all its conversation content as memory is a simple and blunt solution. A better approach is to let the AI manage its own memory. The langmem SDK launched by Langchain some time ago embodies this idea, providing the AI with a set of tools to manage memory, allowing the AI to decide what should be recorded.

When designing Nuwa, this was also the approach taken, providing a set of memory Actions: add/update/remove/compact. Each time an interaction occurs, the AI can call the corresponding Action to maintain its memory. In most scenarios, it can also work, for example, an Agent that distributes test coins to users, limiting each user to only receiving once per day; it will use memory to keep track of the receiving records.

The way this memory works is largely an automatic analysis, evaluation, and summarization of conversations, which still differs from the way real human memory functions.

Does AI really understand "memory"?

A simple test case is to play a guessing game with AI, letting it think of a number, and then you guess. In reality, the AI doesn't actually "think" of a number and then let you guess; it tricks you into interacting with it a few times and pretends you guessed correctly, because it doesn't have a place to store the "thought". Once it has a memory tool, I imagine it would use the memory tool to store the "thought" content without saying it out loud, but in reality, the AI does not naturally understand the relationship between "thinking" and memory. Unless you explicitly tell it, "Please think of a number and save it with the memory tool," it is still just making things up.

This example seems simple, but it actually exposes a key issue: AI at its current stage cannot naturally connect "internal thinking" and "memory." Its use of "memory" is more about responding to instructions rather than evolving proactively.

Memory Challenge in Multiplayer Interaction

A greater challenge arises when placing AI Agents in social environments. How should memory be managed when they interact with multiple people?

If the memory of the AI Agent only spans multiple conversations of a single person, the mechanism outlined above can generally be applied. However, if it exists within a social network and interacts with multiple different users, it will encounter two typical problems:

1. Memory storage and isolation issues: If all interactions of everyone are recorded, loading them every time will easily lead to context explosion.
2. The determination of shared memory: What kind of information needs to be shared across entities? What should be retained in the "memory of a specific user"? This is a judgment that current AI finds difficult to make.

Nuwa's design isolates shared content based on the address of the Agent's interaction object, storing it in the Agent's own address memory. However, this mechanism requires the AI to be aware that "this information is shared," and the practical results indicate that the AI performs poorly.

For example: I transferred a Coin to the AI Agent and told it, "When another user xxx comes to communicate with you, transfer it to him as well." This is a very typical shared memory. But the AI does not understand that this information is its own "commitment" that needs to be saved as shared memory for future use.

The Risks of Memory and Future Directions

The memory capability of the AI Agent still has a lot of room for development. On one hand, it comes from the continuous refinement of prompts and tools by Agent developers, and on the other hand, it relies on the evolution of the model itself. Especially:

1. The ability of memory attribution: Can AI understand whether a piece of information is "my commitment to someone," "someone's request," or "my past speculation"? Currently, this type of "semantic attribution" is still very weak.
2. The relationship between memory and prediction: Good memory is not only about recalling but also about foresight. The information that may be used in the future is actually a form of reasoning about the future.

Memory and State

The memory capability of AI agents still has a long way to go. It is not just a storage issue, but a problem of cognitive structure— it needs to understand what to remember, where to store it, and when to forget.

In fact, we can look at this issue from a different angle. If we understand Prompt as "rules" and memory as "state", then the entire behavior process of the AI Agent is essentially a stateful reasoning system.

From this perspective, the memory interface should not just be a simple ability to "record conversations," but should support a set of structured state types. For example:

1. Users prefer this Key-Value state
2. Historical interaction such a time series
3. Map structure of object state
4. Even more complex graph structures to express social relationships, task dependencies, or causal chains.

Summary

This direction, whether viewed from the perspective of products, algorithms, or system design, is a rapidly evolving and opportunity-filled frontier.