Real Memory summary for on-going conversations (avoid LLM size limits)

[Isaac4real]

Is your feature request related to a problem? Please describe.
the framework has basic truncation and message removal, but NO built-in summarization to manage long-term memory growth. It will eventually break if the memory exceeds the LLM's context window. This is a significant area for improvement.

Describe the solution you'd like
Memory management capable of creating long-term memories (convert older messages into a short summary/facts version). Limit the memory to only last e.g. 100 memories. For older memories, use RAG to retrieve any relevant memory if at all.

Is this not possible with the current options.
this kind of memory management is not possible natively in smolagents

Describe alternatives you've considered
I've developed my own external memory management for long-term memories. but it's a pain to keep it in sync with the main repo version

Additional context
somehow similar issue here: #531

[wgabrys88]

Yeah, another must have feature to even consider smolagents to be usefull...

In theory seems easy, just output in the first iteration a summarization of what was already tried and make sure that the printout is easily noticeable by "second round" - also working on this, maybe we can sync?

BTW, share you solution, maybe it is not you who has to follow to be in sync, it may be the other way ;)