Memory Class Refactoring

[BhoomiAgrawal12]

While working with mesa-llm, I noticed that LLMAgent.__init__ unconditionally creates an STLTMemory instance and there is currently no way to pass a different memory class during construction. Users who want EpisodicMemory, ShortTermMemory or LongTermMemory have to manually reassign agent.memory after creation, which completely bypasses llm_model forwarding.

There is also no hook to customize memory per agent, which I feel becomes a blocker for LLMAgent.create_agents() as well.

The four memory types are designed for different use cases:

STLTMemory : General-purpose agents with balanced memory + cost
ShortTermMemory : Lightweight agents and no consolidation needed
LongTermMemory : Agents needing a summarized view of long history
EpisodicMemory : Importance-weighted recall

However, the choice is currently hardcoded. My understanding is that allowing the memory implementation to be configurable during agent construction would make the design more flexible.

One possible approach could be introducing a memory_class parameter (defaulting to STLTMemory for backward compatibility) along with a memory_kwargs dictionary to forward additional arguments to whichever memory class is selected.

From my perspective this seems like a reasonable gap in the current implementation rather than an intentional restriction, but I wanted to confirm with the maintainers before implementing anything.

If this direction aligns with the project’s design goals, I’d be happy to work on a PR for it.

@jackiekazil @colinfrisch @wang-boyu