What's this about?
Two papers at 9 upvotes each. EM²Mem builds event-centric multimodal memory for LLMs — organising memory around events rather than documents or turns. It Takes Two to Match co-evolves a generative retriever with reinforcement learning.
Why organise memory around events?
Because that is how the thing being remembered is actually structured. A conversation is not a list of turns any more than a week is a list of hours — it is a sequence of things that happened, each with participants, a time, a place and consequences. Chunking by turn or by document splits single events across boundaries and merges unrelated ones, which is why retrieval so often returns something adjacent to what you needed. Event-centric memory tries to make the retrieval unit match the unit of meaning.
Why is it worth noticing?
- It is the structured counterpoint to Memoryfields' "just use markdown files" argument from two days ago, and both are defensible.
- Co-evolving the retriever with the generator addresses the standard RAG failure: a retriever optimised in isolation for a generator it never sees.
- Both are early and lightly upvoted; noted for the direction.