Adding Memory Components (useful for RAG workflows) in AGS

[victordibia]

Add the ability to attach memory to agents in AGS.

What

A common pattern (or requirement) for agentic apps is to support the ability to add memory to agents.
Here memory refers to the ability to augment model context based on the state of the agent (previous message).

It is valuable to support this in a tool like AGS. An example would be that users can:

• Define a Memory component (similar to how tools, models, etc can be defined today). Memory might be a simple list backed by a file on disc, or a vector
• Attach the memory to an agent .. similar to how you can drag a tool into an agent specicifcation
• At runtime, the agent will query that memory store and update context before making an LLM call.

How

• Support Memory as a high level component in AgentChat. Specifically for the AssistantAgent. This work is In progress [Draft, Feedback Needed] Memory in AgentChat  #4438
• Enable configuration serialization protocol for AgentChat agents. This will make it easy to convert a memory component to a json dict configuration.
• Design a set of example memory components (provider classes etc) with support for providers likeChromaDB, PineCone, PGVector, MongoDb etc These classes will likely live in 3rd party libraries and developers can implement their own.
• Update AGS UI to support memory components
  • Optional - add UX to interact with some memory config .. load the config, list content, add to it, edit it etc ... since this will vary with implementation, some work is needed to standardize.
• Test, test test ....

This issue is quite large and might take a while, but should address the most common use cases.
cc @Knucklessg1