[Feature Request]: Agent supports global memory

[TeslaZY]

Is there an existing issue for the same feature request?

• I have checked the existing issues.

Is your feature request related to a problem?

If Agent uses the Category node, it will sometimes divide the user's questions into different categories, but in fact, the two problems are contextual. For example, the first question requires agent to recommend a product. After interacting with llm, llm prompts you to need more information, such as the usage scenario of the product. At this time, if I directly reply to the use scenario name, Category will divide the scene name and the first question "recommend a product" into two categories. Then the agent will follow different flow branches and make memory switching, which is different from the previous tag.

Describe the feature you'd like

We can consider adding the global agent memory to support better problem classification, question answer and other scenario performance, and avoid the context conversion problems caused by the switching branch caused by Category. Agents should always behave like the same person for a session, instead of looking like schizophrenic now.

Describe implementation you've considered

No response

Documentation, adoption, use case

- Add the global agent memory 

Additional information

No response

[KevinHuSh]

We typically use multiple categorizes which represents like a tree. The scene name is on sub-tree of the recommend a product node .

[TeslaZY]

In some cases, it is not suitable to use sub-tree.

[I hope the agent can divide the two problems into one category according to the context]

User: Recommend me a GPU

Agent: Please tell me the GPU usage scenario you need.

User: machine learning

Agent: return {the product information of machine learning GPU from the knowledge base or database}

[I hope it is divided into two categories]

User: machine learning

Agent:return {Information from the knowledge base or search engine}

Therefore, agent memory and global context may be necessary.