WikiSkill: Compiling Agent Experience into Persistent Knowledge

WikiSkill compiles agent trajectories into persistent, queryable knowledge, allowing systems to reuse successful strategies without retraining. The method treats past interactions as a skill database rather than discarding them after task completion.

This shifts the operational bottleneck from model inference to knowledge management. For builders, it means agents no longer need to rediscover solutions for recurring tasks, reducing token spend and latency on routine operations. The infrastructure requirement changes: you now need a storage and retrieval layer for behavioral data, not just model weights. This makes long-horizon autonomy more feasible, as agents can accumulate competence over time without intervention. The second-order effect is a workflow change—evaluation shifts from static benchmarks to measuring how quickly an agent converts new experiences into reusable skills. Builders who ignore persistent knowledge risk deploying systems that plateau at their initial training distribution, while competitors compound learning with each interaction.