Boost AI Workflow: Personal Knowledge Base with llm_wiki and Evolving Agents

Retrieval‑augmented generation (RAG) has long suffered from a fundamental flaw: after each interaction the AI starts from scratch, forgetting any previously processed documents or insights.

llm_wiki: Turning Documents into a Persistent Wiki

Inspired by a design pattern proposed by Andrej Karpathy, llm_wiki is a cross‑platform desktop application written in TypeScript that "compiles" uploaded PDFs, web pages, and articles into a durable wiki rather than merely indexing them for one‑off retrieval.

Core logic : The app applies a two‑step chain‑of‑thought process—first analyzing the document, then generating a wiki page that includes source provenance and caching. New material is incorporated incrementally into an existing knowledge graph instead of rebuilding it.

The knowledge graph maintains four signal types—direct citations, source overlap, Adamic‑Adar similarity, and type affinity—and automatically discovers clusters using the Louvain community‑detection algorithm.

Deep Research mode : One‑click multi‑source web search; results are auto‑ingested, allowing the knowledge base to grow organically.

Chrome Web Clipper plugin : Capture web pages with a single click and queue them for processing.

Async Review System : LLM‑flagged items that need human judgment are pre‑generated with retrieval suggestions, enabling batch review at a convenient time.

Obsidian compatibility : The wiki directory can be opened directly as an Obsidian vault, integrating with everyday note‑taking workflows.

The project positions itself as "not a search engine, but a growing personal knowledge base" suitable for researchers, heavy information processors, and anyone with large document collections.

https://github.com/nashsu/llm_wiki

SkillClaw: Evolving Skills for AI Agents

While llm_wiki addresses human knowledge accumulation, SkillClaw tackles the neglected problem of AI agents losing their experience after each session. Most agent frameworks reset after a conversation, discarding useful operations, pitfalls, and discovered tricks.

SkillClaw extracts reusable Skill.md entries from real sessions and stores them in shared storage, making the skills available to all agent instances.

Three‑layer architecture

Client Proxy : A local API proxy intercepts /v1/chat/completions requests, records session data, and synchronizes the skill library.

Workflow Evolve Server : Executes a fixed three‑step pipeline—Summarize → Aggregate → Execute—to distill generic skills from session logs.

Agent Evolve Server : Allows agents to decide autonomously how to refine skills, offering greater flexibility for complex scenarios.

All three modules share the same storage backend (supporting Alibaba OSS, S3, or local filesystem) and use a unified SKILL.md format that can be mixed and hot‑swapped.

Plug‑and‑play : Install with two commands— skillclaw setup && skillclaw start —without modifying existing agent code.

Cross‑Agent sharing : Skills discovered by one user are instantly available to the whole group.

WildClawBench validation : Benchmarks on public datasets show that Qwen3‑Max gains significant performance improvements in limited‑group interactions solely by accumulating experience, without changing the underlying model.

The system is compatible with multiple agent frameworks and supports code generation, content creation, and tool‑calling use cases.

https://github.com/AMAP-ML/SkillClaw

Shared Belief

Both projects share the conviction that the biggest waste in current AI systems is the loss of knowledge after each dialogue. By turning transient document lookups into a structured, growing knowledge base and converting fleeting agent actions into reusable skills, they aim to move AI from a mere "library clerk" to a system that truly remembers and learns.

If you find these tools useful, feel free to share them with others building AI workflows.