DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
The paper claims agents can build useful memory by practicing on tasks they generate themselves.
DAEDALUS uses an explorer agent to create solvable challenges and a solver agent to attempt them. Failures are turned into heuristics, but only kept after repeated success with that heuristic in context. The accepted heuristics become a memory bank used at test time. The authors report gains of up to 15.9 mean success-rate points and up to 2.2x pass^5 over a no-memory baseline across AppWorld, τ²-bench, and AutomationBench. ArXiv · AI/CL/LG's note
DAEDALUS uses an explorer agent to create solvable challenges and a solver agent to attempt them. Failures are turned into heuristics, but only kept after repeated success with that heuristic in context. The accepted heuristics become a memory bank used at test time. The authors report gains of up to 15.9 mean success-rate points and up to 2.2x pass^5 over a no-memory baseline across AppWorld, τ²-bench, and AutomationBench. ArXiv · AI/CL/LG's note
score 6