Useful memory changes what reasoning can reach.
Real work unfolds over time. Intelligent systems need to recognize what matters, preserve it with context, and bring it back when a decision depends on it.
The central question.
What belongs in model weights, and what belongs in external memory?
Questions that follow
- How can systems keep source evidence distinct from inference?
- When should information be stored, refreshed, or forgotten?
How we study it
This direction examines retrieval, context compression, provenance, and tasks that revisit earlier work. The focus is useful recall: relevant information, clear sources, and enough context to judge whether it still applies.
What meaningful progress looks like
Look at retrieval precision, source fidelity, recall over time, and whether remembered context improves decisions without carrying stale assumptions forward.
Explore all research programsRead the work behind these questions.
These independent papers and official guidance offer useful starting points for this field. Follow each link to review the original methods, evidence, and limits in context.