DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks

DAEDALUS. Bootstrapping Agent Memory from Self-Generated Tasks. By Antoine Edy and five others, from Illuin Technology. Published on the sixth of October. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Put an LLM agent in a new environment and it often lacks the operational knowledge to act reliably. It has to discover for itself how specific tools behave and what conventions the environment follows. Without memory of earlier attempts, it repeats the same mistakes from one task to the next, which means more failed tasks and longer trajectories. The usual remedies each assume something. Some systems rely on guidelines written by humans. Others build procedural memory from training tasks checked by an oracle verifier. Both require prior knowledge of the environment, which is precisely what a new environment does not come with.

DAEDALUS bootstraps reusable agent memory from practice the agent generates itself, with no existing tasks and no oracle verifiers. It pairs two agents. An explorer interacts with the environment and generates tasks that are challenging yet solvable. A solver attempts them. The environment becomes its own curriculum.

Learning happens at the solver's failures. From each failure a heuristic is derived, and that heuristic is accepted only after the solver, with the heuristic in its context, succeeds repeatedly. A lesson has to prove itself before it is kept. The same outcomes feed back to the explorer, which uses them to refine the difficulty of the tasks it generates next. Accepted heuristics are finally consolidated into a memory bank that the agent uses at test time.

The pieces form a loop that checks itself. The explorer needs the solver to know whether a task was too easy or too hard. The solver needs the explorer to supply tasks hard enough to fail at. And the acceptance test stands in for the oracle verifier that other methods require: instead of an outside judge declaring an answer correct, a heuristic earns its place by turning failure into repeated success. Nothing in this loop needs a human-written guideline or a set of training tasks prepared in advance.

The evaluation spans AppWorld, tau squared bench, and AutomationBench. Against a baseline with no memory, DAEDALUS improves mean success rates by up to fifteen point nine points, and improves pass to the fifth by up to two point two times. It is competitive with methods that use training tasks, and its inference cost is lower than most of them. Gains already appear with a small exploration budget, and the heuristics it learns also help agents from other model families.

The ablations say where the value comes from. Traces from the solver supply the key information needed to derive effective heuristics. Factorising early discoveries makes exploration more cost-efficient. And beyond building memory, the tasks DAEDALUS generates can serve as a proxy for benchmark tasks when ranking models by performance.

The design trades a dependency for a discipline. Instead of needing someone who already knows the environment, whether a human writing guidelines or a verifier judging training tasks, DAEDALUS asks the agent to invent its own practice and keeps only the lessons that measurably help. That the practice tasks can also rank models suggests they capture something real about the environment, not just about the agent that made them. Code and artifacts are publicly available.