Agent memory has become its own subfield, and the stress-test papers arrived with it
A cluster of recent work treats memory as the central problem of long-horizon agents rather than a feature of context length. MemFail stress-tests failure modes directly; EMBER budgets evidence retention; RaMem does contextual reinstatement for long-term recall.
The reframing is the substance. Context length is a capacity question and memory is a policy question — what to keep, what to discard, what to reconstruct and when. A million-token window does not answer any of those, which is why the work continued after the windows got large.
MemFail is the healthiest sign in the cluster. A subfield that produces a dedicated failure-mode stress test has moved past demonstrating that its systems work and into finding out where they break, which is the transition that separates a technique from a discipline.
EMBER's budgeted evidence retention names the real constraint. Retention is not free, so the design problem is an allocation problem under a budget, and that framing makes the work engineering rather than architecture aesthetics.
arXiv — MemFail: stress-testing failure modes of LLM memory systems → · arXiv — EMBER: efficient memory via budgeted evidence retention for long-horizon agents → · arXiv — RaMem: contextual reinstatement for long-term agentic memory →