DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
DAEDALUS builds agent memory by making agents practice on tasks they generate themselves.
The paper pairs an explorer agent that creates difficult but solvable environment tasks with a solver agent that attempts them. Failed solver traces are turned into heuristics, which are kept only if they repeatedly help the solver succeed. Those accepted heuristics become a memory bank for later use. The authors report gains of up to 15.9 points in mean success rate and up to 2.2x in pass^5 over a no-memory baseline. HF Daily Papers' note
The paper pairs an explorer agent that creates difficult but solvable environment tasks with a solver agent that attempts them. Failed solver traces are turned into heuristics, which are kept only if they repeatedly help the solver succeed. Those accepted heuristics become a memory bank for later use. The authors report gains of up to 15.9 points in mean success rate and up to 2.2x in pass^5 over a no-memory baseline. HF Daily Papers' note
score 5