Defeasible Library Learning: Typed Methods with Runtime Contracts that Un-learn on Drift

We present defeasible library learning: an LLM agent reuses past work as a learned library of typed, composablemethods, each carrying a defeasible runtime contract. Where retrieval (RAG), case-based reasoning, and prose skilllibraries recall by surface similarity — and so keep serving a stale answer when a premise silently becomes false —a method here is assumed when composed, asserted when run, and retracted with blame when its guarantee fails (a JTMStruth-maintenance step), after which the library is surgically revised rather than discarded. The same typed structure makes reuse cheap (recurrent cases skip the model call) and composition checkable oncontracts alone, at bounded per-call context. On a real rule-driven engine (a deterministic stub plus a live localmodel), recall-only memories go stale under mid-stream drift while the typed contract recovers correctnessselectively — for a single method, across a learned method chain, and head-to-head against MemGPT, Reflexion, andGraphRAG. Every mechanism is prior art; the contribution is their composition into one representation that un-learnson drift, bounded by a measured canonicalizability ceiling. Full abstract, experiments, and references are in the PDF. Code and the self-contained reproducibility artifact arepublic at github.com/9pings/skynet-graph.