Google Research: Knowledge Wikis for Evolving AI Agent Skills

X postDAIR.AIResearch

A Google research paper presents an architecture for evolving agent skill libraries by separating raw execution traces, a persistent knowledge wiki, and executable skills.

Key points

  1. Separates three traditionally merged layers: raw execution traces, a persistent wiki of accumulated knowledge, and executable skills.
  2. Agent experiences consolidate into the central wiki, forming the foundation for future skill updates rather than relying on fragmented optimization logs.
  3. Ablation studies confirm that the persistent wiki architecture accounts for most of the performance gains.
  4. Smaller models utilizing evolved skills outperform significantly larger models that lack them.
  5. Evolved skills transfer across model families, with externally developed skills occasionally outperforming self-evolved alternatives.

Why it matters

Managing skill libraries is critical for production AI agents. By decoupling accumulated knowledge from executable code, engineers can achieve superior performance with smaller, cheaper models and seamlessly share capabilities across different model architectures.

Analysis

Skill evolution mechanisms often suffer from context bloat and brittle updates because they conflate raw session logs with skill logic. Google's decoupled wiki approach creates a persistent memory layer that stabilizes skill updates. The finding that cross-family skill transfer can outperform native skill generation offers a practical blueprint for distilling agent behaviors from frontier models to lightweight deployments.

Original post

DAIR.AI

@dair_ai · 28 August 2026

Banger paper from Google. If you maintain a skill library for your agents, you might want to check this out. (bookmark it) This work separates three things that skill-evolution systems usually collapse into one. Raw execution traces, a persistent wiki of accumulated knowledge, t.co/HPi2FFIbJM

4k likes87 replies8.6k bookmarks500.7k views

View on X
Selected replies (5)
  • @CaryPalmerr @dair_ai Yes! This is why I built t.co/C4RGpNZdDq a while back. Graph retrieval of tools/skills/rules injected into turns. Way better than local .MD bloat. You can also carry your context between any harness or machine, share context with others via qr, etc t.co/rniS8cuW5C
  • @alpniks @dair_ai This looks a lot like ReasoningBank, an earlier Google paper from this year: t.co/nNWBbzBc0K
  • @ovitrif @dair_ai If this is gonna improve my skills as great as they improved Gemini then I have all the reasons to ignore it. Usually, in this world, you have authority on a subject when your results proved you're worth listening to. Google so far is still trying.
  • @gmccreight @dair_ai Amazing! I'm working on t.co/ZKlKh27wkt for curating and sharing subgraphs of personal context. I believe skills are just a subset of important context. It's open source, but also offers hosting the bundles as microsites.
  • @iav @dair_ai The knowledge used to improve a process does not necessarily belong inside the process itself 🤔

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu

    WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

    This research paper introduces WikiSkill, a framework that co-evolves AI agent skills alongside a persistent knowledge base (wiki). By disentangling raw execution experience, accumulated knowledge, and executable skills, WikiSkill enables systematic reuse of insights across iterations.

    AI & LLMs· Research

  • Article:Hongjun Liu, Yifei Ming, Shafiq Joty, Chen Zhao

    HASP: Harnessing LLM Agents with Executable Skill Programs

    A research team introduces HASP (Harnessing LLM Agents with Skill Programs), a framework that converts past agent experiences into executable Program Functions (PFs). Instead of relying on passive textual prompts, PFs act as executable guardrails that trigger on failure-prone states to alter the next action or supply corrective context.

    AI & LLMs· Research

  • Article:Siru Ouyang, Jun Yan, I-Hung Hsu, et al.

    ReasoningBank: Scaling Agent Self-Evolution with Reasoning Memory

    ReasoningBank is a memory framework for LLM agents that distils generalisable reasoning strategies from self-judged successful and failed experiences, enabling continuous self-evolution over time.

    AI & LLMs· Research

  • Video

    Video:AI Engineer

    Don't Ship Skills Without Evals: Best Practices from Google DeepMind

    Google DeepMind's Philipp Schmid explains why AI agent skills require systematic evaluation suites and how to build lightweight eval harnesses. Drawing on data from SkillsBench and internal practices at Google, he outlines how to write, structure, test, and eventually deprecate skills as foundation models improve.

    AI & LLMs· Talk

  • Link:Anthropic

    Agent Skills: An Open Standard for Modular AI Agent Capabilities

    Agent Skills is an open format designed to equip AI agents with specialised domain knowledge and repeatable workflows. Centred around a directory containing a SKILL.md file, it relies on progressive disclosure to keep context overhead minimal.

    AI & LLMs· Resource

  • Link:agentwikis.com

    Agent Wikis: Maintained Knowledge Bases for Humans and AI Agents

    Agent Wikis transforms fast-moving information sources such as repositories, changelogs, transcripts, and community Q&A into structured Markdown knowledge bases. Designed for both humans and AI agents, the content is accessible via an llms.txt entry point.

    AI & LLMs· Resource

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.