What's this about?
Metis proposes a foundation model for memory — treating an agent's long-term memory as a trainable component in its own right. It's part of a cluster of memory papers today (also "Memory for LLMs" and "MemHarness").
Why it matters
Memory is a recurring failure point for agents (6 upvotes); a wave of papers treating it as a first-class, trainable system — rather than a bolted-on vector store — signals where a lot of agent research is heading.