What would be an approach to implement basic memory, and have the agent act differently based on that memory?

[gkpo]

Question

I have an agent with several tools and I want it to have basic memory - stored persistently (e.g., in a database) - so it can decide whether to call tools based on stored values.

For example, the memory might contain key - value pairs like:

userId: 543

userName: John

bookingNumber: 36dbsb-jd63dh-77fghds-abxhdg64-cbsj745js

I want the agent to make decisions based on this stored state. From what I’ve read, the typical approach is to use a dynamic master prompt that changes according to the memory object.

For example:

"You are a helpful assistant. If the username is not defined, ask for it and then store it using the store_username tool. The current user name is ---."

However, I don’t like this method because it relies on the agent holding mutable data in the prompt, which could be corrupted or altered during execution. For instance, I think the booking number is more error-prone than the user ID.

Does my goal make sense? Any ideas or advice would be appreciated.

Additional Context

No response

[DouweM]

However, I don’t like this method because it relies on the agent holding mutable data in the prompt, which could be corrupted or altered during execution. For instance, I think the booking number is more error-prone than the user ID.

@gkpo Can you help me understand what you're worried about getting corrupted/altered?

The typical approach is to give the model an add_memory tool, and then either search_memory+get_all_memories tools, or automatically putting all memories in the instructions, as you mentioned. The instructions function is reevaluated at every step of the agent loop, so the model would always have fresh information.

Often memories can be any text, so a search tool would be needed, but in your case you could have a store_memory tool with key: Literal['userId', 'userName', 'bookingNumber'], value: str args, so that the model is restricted to those keys, and then a load_memory tool taking that same key, which returns a single value directly. Or you could have a get_all_memories tool that the model can call to get a dict with all keys and values, as an alternative to proactively putting them into the instructions.

[AlexEnrique]

I'm trying to answer the same question about memory. I am considering give the model a tool as @DouweM said and add the information via a dynamic instructions (some models "seems to take system prompts more seriously" than the return of a tool). I wonder if there is a better approach.

In my case, the agent is generic, customized by the user, so that it may add any kind of information in the memory, e.g. an address, a phone number, or any other information the user may instruct.

To change the behavior, you may add or remove an instruction based on a memory value, e.g. the user name. I've done this with good results.

[gkpo]

@DouweM thanks for your recommendations.
So to answer your question, when i say some data could be corrupted, i'm talking about hallucinations.
If the instructions contain some short string or id ("the client id is 123") that's an easy thing to pick up by the agent to use in tool parameters. But if the string is longer (like the booking number i gave in my example: "36dbsb-jd63dh-77fghds-abxhdg64-cbsj745js"), then i guess it would be easier for the agent to make a mistake in one character.

I'm interested in the 3 tools approach you mentioned.
Just to clarify, could you describe what each of these tools do?
add_memory is pretty simple i guess
search_memory would this be equivalent to get_memory with a fixed and rigid key ? or something more flexible like vector search or string similarity ?
list_all_memories why would we need this tool? and what would it list? only the keys? or the keys and the values?

Thanks in advance!

[DouweM]

@gkpo Hallucinations are not much of an issue in recent models, and they're very good at repeating themselves or other text in the context, so I wouldn't worry about it messing up the ID.

search_memory would this be equivalent to get_memory with a fixed and rigid key ? or something more flexible like vector search or string similarity ?

Yeah that'd typically be a similarity search

list_all_memories why would we need this tool? and what would it list? only the keys? or the keys and the values?

It'd print all of the keys and values for cases where the model is asked something like "what do you know about me?". But it's definitely not necessary to have.

[github-actions]

This issue is stale, and will be closed in 3 days if no reply is received.

[github-actions]

Closing this issue as it has been inactive for 10 days.