Releases·Tool Release
WikiSkill implementation released for Hermes Agent: self-evolving agent skills via persistent knowledge wiki
A GitHub project implements WikiSkill (arXiv:2608.27454), compiling agent experience into a persistent knowledge wiki so skills evolve themselves. The release targets the Hermes Agent and includes real agent runs, isolated skill gating, and a documented live run log. It gives practitioners a production-minded reference implementation of the paper's approach.
42 engagement1 sourceSat, Aug 29, 2026, 08:44 PM
AI Overview
- Faithful implementation of Algorithm 1 from the WikiSkill paper (arXiv:2608.27454), with real agent runs rather than synthetic demos
- Core mechanism: compile agent experience into a persistent knowledge wiki so skills evolve themselves
- Includes isolated skill gating and a documented live run log
- Ships with a docs site (ashutoshsinghpr7.github.io/wikiskill) alongside the GitHub repo
Related
Developments·Paper
Google Research's WikiSkill compiles agent experience into a persistent wiki, letting a 9B model beat a 27B rival
A Google Research preprint introduces WikiSkill, a framework that co-evolves agent skills with a persistent wiki knowledge base instead of relying on raw transcripts. It separates immutable execution traces, consolidated knowledge, and compiled skill files, with a validation gate keeping only score-improving updates. Reported results include a 9B model with evolved skills beating a 27B rival across five agent benchmarks, and skills transferring across model families — though one transferred skill cut Gemini's score from 50.5% to 18.1%.
A Google Research preprint introduces WikiSkill, a framework that co-evolves agent skills with a persistent wiki knowledge base instead of relying on raw transcripts. It separates immutable execution traces, consolidated knowledge, and compiled skill files, with a validation gate keeping only score-improving updates. Reported results include a 9B model with evolved skills beating a 27B rival across five agent benchmarks, and skills transferring across model families — though one transferred skill cut Gemini's score from 50.5% to 18.1%.
Updated Mon, Aug 31, 2026, 01:07 PM — Community discussion highlights a failure case: one transferred skill cut Gemini's score from 50.5% to 18.1%.
11035 engagement3 sourcesarxiv, redditThu, Aug 27, 2026, 05:59 PM
Releases·Tool Release
Hermespace: persistent agent world for Hermes agents released on GitHub
Hermespace is an append-only persistent world for Hermes agents, recording every session, belief, landmark, and evolution without pruning or decay. It provides a pocket workbench with limited FOA desk, dual decode, skills+MEMORY fabric, and neural FOA, enabling agents to build a deepening model of themselves and their environment across sessions.
Hermespace is an append-only persistent world for Hermes agents, recording every session, belief, landmark, and evolution without pruning or decay. It provides a pocket workbench with limited FOA desk, dual decode, skills+MEMORY fabric, and neural FOA, enabling agents to build a deepening model of themselves and their environment across sessions.
69 engagement1 sourcegithubSat, Jul 18, 2026, 01:23 PM
Developments·Paper
Two arXiv papers advance agent skills: SkillAlchemy for open-world skill creation, SPT for skill pre-training
Two August 2026 arXiv papers tackle agent skills from complementary angles. SkillAlchemy (Aug 24) studies open-world skill creation, where a creator agent must discover behavior-relevant requirements omitted from underspecified skill briefs. SPT (Aug 27) introduces Skill Pre-Training, a mid-training method applying causal language modeling to public skills, treating them as pre-training data rather than inference-time context.
Two August 2026 arXiv papers tackle agent skills from complementary angles. SkillAlchemy (Aug 24) studies open-world skill creation, where a creator agent must discover behavior-relevant requirements omitted from underspecified skill briefs. SPT (Aug 27) introduces Skill Pre-Training, a mid-training method applying causal language modeling to public skills, treating them as pre-training data rather than inference-time context.
0 engagement2 sourcesarxivMon, Aug 24, 2026, 03:58 PM
Developments·Paper
VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience
Researchers propose VCE-Skill, a method that distills evolution knowledge from public skill version histories to improve agent skill self-evolution. It complements existing trajectory-based revision by providing reusable priors from version changes, addressing a gap in current approaches.
Researchers propose VCE-Skill, a method that distills evolution knowledge from public skill version histories to improve agent skill self-evolution. It complements existing trajectory-based revision by providing reusable priors from version changes, addressing a gap in current approaches.
0 engagement1 sourcearxivMon, Aug 17, 2026, 01:16 PM
Products·launched
AI-Agent-Skills: open-source knowledge base for building AI agents
AI-Agent-Skills is an open-source, structured knowledge base that compiles notes on building AI agents into 20 categories, covering reasoning strategies, tool use, RAG, MCP, multi-agent patterns, safety, and deployment. It serves as a comprehensive handbook for developers, combining the organization of Awesome Lists with the depth of Papers with Code, aiming to replace scattered blog posts and research rabbit holes.
AI-Agent-Skills is an open-source, structured knowledge base that compiles notes on building AI agents into 20 categories, covering reasoning strategies, tool use, RAG, MCP, multi-agent patterns, safety, and deployment. It serves as a comprehensive handbook for developers, combining the organization of Awesome Lists with the depth of Papers with Code, aiming to replace scattered blog posts and research rabbit holes.
1 engagement1 sourceredditSat, Aug 1, 2026, 11:46 AM