help: how to write a custom memory layer?

[xieydd]

I want to use this project to implement my agent framework, but I want to rewrite my memory layer.Any suggestions? @evalstate

[evalstate]

Hey @xieydd -- would be good to hear a little more about your needs before making suggestions.

There is a slightly related discussion here: #119

Memory management is all handled in augmented_llm.py. There is a Memory class for handling provider specific types, and message_history for the MCP compatible types. Note there is special handling of applied prompts, and the Agent configuration option to use history or not.

LMK if you need more guidance, or have any ideas to make this more extensible (pr etc.)

[xieydd]

@evalstate Thanks for your quick response. I'll describe the scene in detail.

We have a RAG service that provides semantic search and full-text search. In the next release, we will also provide graph indexing after entity extraction. Fast-agnet has elegantly implemented the LLM, Tool module, and the above-mentioned session-level Memory in the Agent. However, we would like to have access to more types of Memory, such as Long-Term Memory or Entity Memory. I don't want to modify fast-agent's code to implement the above scenario, can fast-agent provide such an abstraction and an interface for the implementation of the specific implementations.

[xieydd]

@evalstate memory layer like mem0.

[xieydd]

@evalstate Any ideas?