Researchers have proposed a framework called WikiSkill that pairs an AI agent’s reusable skills with a persistent, continuously updated wiki of accumulated experience. The paper appeared on arXiv on August 27, 2026, credited to Liyan Tang and coauthors. The abstract does not state an institutional affiliation.
The core idea addresses a specific failure mode in agent systems. Prior methods let agents discover skills automatically from their own trial and error, but the reasoning that produced each skill usually gets thrown away once the skill is saved. WikiSkill keeps that reasoning. It separates three layers: raw execution traces, a consolidated knowledge base, and the executable skills themselves, then feeds new experience back into the knowledge layer so later skill updates can draw on it.
That structure matters because most agents deployed in production today do not remember anything between sessions. Ask a coding agent to fix a build error on Monday and it may hit the identical error on Friday with no memory of the fix, because most skill or memory systems capture the output of a session without capturing why it worked. Teams running agents at scale cite this gap as one of their most common complaints: the tooling improves, but the agent itself accumulates no judgment.
According to the abstract, the authors tested WikiSkill across multiple benchmarks and models and found it consistently outperformed existing skill-evolution methods, as well as agents given no skills at all, in most model-benchmark combinations. Two findings stand out. Skill evolution and raw model scale appear to be complementary rather than substitutes. Larger models tend to gain more from evolved skills, and smaller models equipped with skills can beat larger models that lack them.
Skills also transfer across models and model families, per the abstract. In some cases, skills evolved by one model outperformed skills a model evolved for itself, which points toward shared skill libraries rather than model-locked ones.
The authors’ ablation work isolates the wiki as the load-bearing piece of the system. Removing the persistent knowledge accumulation step degrades skill evolution, which suggests the gains come from consolidating and reusing experience rather than from the skill-authoring mechanism alone.
The paper’s abstract does not report specific benchmark scores, model names, or dataset sizes, so those details await the full text or a later revision.
For teams building agent products, the finding worth tracking is the cross-model transfer result. If skills genuinely generalize across model families, that weakens the case for building memory systems tightly coupled to one vendor’s model. Treat accumulated agent knowledge as portable infrastructure to evaluate now, not a model-specific asset to wait on.
Findings per the abstract of “WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution,” posted to arXiv on August 27, 2026.
