Evo-Memory

Fits agents that are expected to get better over a deployment by learning from their own past runs. Borrow the idea of turning a fixed test set into a sequential stream and measuring gain over time.

What it asks the agent to do

• solve a stream of related tasks and reuse what earlier tasks taught
• update stored experience after each interaction
• answer single-turn reasoning questions using memory of past attempts

What to borrow from it

grading designbaseline

Caveats

Code was promised but no repository link was found; the ten source datasets are existing public sets re-ordered, so leakage from pretraining is possible.

Details

Links

• Paper