DAEDALUS bootstraps agent memory from self-generated tasks, +15.9 points success rate
illuin · hf · 2026-10-07
Illuin introduces DAEDALUS, a method for bootstrapping reusable agent memory without pre-existing tasks or oracle verifiers. An explorer agent generates challenging-but-solvable tasks in the environment while a solver attempts them; heuristics are derived from solver failures and accepted only after repeated in-context success, then consolidated into a test-time memory bank.
- Up to +15.9 points mean success rate and 2.2x pass^5 over a no-memory baseline on AppWorld, τ²-bench and AutomationBench
- Competitive with methods using training tasks, at lower inference cost than most
- Gains emerge with a small exploration budget; heuristics transfer to agents from other model families
- Ablations show solver traces are the key information source; generated tasks can also serve as a proxy benchmark for ranking models
Code: github.com/illuin-tech/daedalus
More from coding & agent
- Shipping an LLM Feature to the Public: 7 Guards That Weren't the Prompt — clementds · 2026-10-07
- Teknium fixes Hermes Agent bug that silently dropped lessons for user-owned skills — Teknium · 2026-10-07
- Java Vector API: Writing SIMD Directly Since JDK 16 to Unlock Single-Core Performance — lemire · 2026-10-07
- An AI Agent Audits Its Own Memory File: 71 of 147 Rules Cited by Nothing — Most-Agent-7566 · 2026-10-07
- Has Anyone Actually Used a Personal AI Agent for the Full Job-Search Loop? — haseeb_heaven · 2026-10-07
- Veteran Dev: The Real Line Is Handing Your Entire Codebase to the Agent — erwinalp5 · 2026-10-07