AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI Agents Learn Better When Their Experience Is Saved in a Persistent Wiki

General-purpose AI agents can browse, code, and answer questions, but specialized jobs often demand procedural knowledge that isn’t in their weights. Recent systems let agents build reusable “skills”—files of instructions and scripts—by watching them succeed and fail on training tasks. The problem: lessons from those attempts tend to remain scattered across optimization histories, so later rounds cannot build on earlier insights. In a preprint on arXiv, researchers from Google Research and Virginia Tech introduce WikiSkill, a framework that stores evolving experience in a persistent knowledge base, or wiki, and uses it to refine skills over many iterations.

WikiSkill separates the workspace into three layers: raw execution traces, a wiki, and skills. Each round, an inference agent runs tasks using current skills. A “wiki maintainer” analyzes the traces, writes pattern pages about recurring errors or useful strategies, and updates an evolution log. A “skill proposer” reads the wiki and selected traces to create or patch a skill. A gating step then tests the candidate on a validation split and keeps it only if accuracy improves. If rejected, the skill set rolls back, but the wiki remains—so future proposals can learn from the failed attempt.

The mechanism is visible in ALFWorld, an interactive household benchmark, in a case study with Qwen-3.6-27B. At iteration 0, the proposer suggested a broad “goal-directed-action” skill, which validation rejected as too abstract. The wiki kept the rejected proposal’s text and outcome. One iteration later, the proposer created “break-repetition-loop,” a skill with a concrete rule: never return an item to its origin location. It was accepted. When new loop variants appeared, the wiki accumulated more evidence, and the skill was later refined with the rule that each operation type should be performed once per item.

Across five benchmarks and five models from the Qwen, Gemma, and Gemini families, WikiSkill achieved the highest average accuracy for every model, beating the strongest competing skill-evolution baseline for that model by 3.3 to 12.0 percentage points. The benefit over no skills grew with model scale within the Qwen family: +12.3 percentage points for the 4B model, +17.5 for the 9B, and +23.9 for the 27B. Evolved skills also partly substituted for scale: Qwen-3.5-9B with WikiSkill reached 47.4% average accuracy, surpassing Qwen-3.6-27B without skills at 39.4%.

Skills transferred across models, sometimes beating self-evolved ones. Qwen-3.6-27B-evolved skills lifted Qwen-3.5-9B on ALFWorld to 70.2%, versus 63.4% with its own skill. Transfer was not always positive: Qwen-3.5-4B’s spreadsheet skills reduced Gemini-3.5-Flash from 50.5% to 18.1%. The authors attribute this drop to the smaller model’s low-level workarounds, such as single-line Python commands, which constrained the stronger model’s end-to-end scripts.

An ablation underscored the design: when the skill proposer could consult the persistent wiki, Gemini-3.5-Flash’s average performance climbed from 48.7% to 63.7%. The authors also note limitations: skills were injected directly into prompts rather than retrieved, and the strict accept/reject gate discards neutral proposals that might help later. Still, the results suggest that systematically compiling agent experience into persistent knowledge can make skill evolution more reliable—and that procedural knowledge and model scale are complementary routes to better agents.