Wiki Memory (LangChain)
Harrison Chase's (LangChain) post (30 Jun 2026) that gives the llm-wiki pattern a name — "wiki memory" — and reframes it as a type of agent memory. T3 (vendor blog, conceptual — "no APIs, benchmarks, or LangChain-specific implementation details," so it's thought-leadership, not a product page). Notable mostly because the framing now comes from the agent-framework establishment, the way OKF did from Google Cloud.
The claim
Wiki memory is "an agent-maintained data structure that transforms raw source data into a compact, persistent, agent-readable knowledge layer." The pattern: "run an agent over some source material and ask it to create a set of files that future agents can use to understand the domain faster." The result is "persistent, structured, inspectable, and updated over time" — plain files, not a database.
Two things it adds to llm-wiki
- A sharpened RAG contrast. "RAG usually retrieves raw chunks at query time. A wiki precomputes and maintains a higher-level synthesis, so the agent does not have to rediscover the structure every time." This is exactly the llm-wiki "compile once, don't re-discover per query" point, stated as a memory-architecture choice against retrieval-augmented-generation.
- A scope boundary (as memory). Wiki memory suits durable domain knowledge — but explicitly not short-term conversation state, user preferences, or high-frequency event logs. Naming what the pattern is not for is the new contribution: it slots the LLM-wiki into an agent-memory taxonomy (bridge to agentic-tooling-wiki's agent-memory) rather than treating it as a standalone PKM idea.
Instances it names
- Karpathy's LLM Wiki — already the founding node here (llm-wiki-gist).
- deepwiki (Cognition) — AI-generated GitHub repository documentation.
- autowiki (Factory) — codebase documentation that stays current.
The last two extend the pattern into a codebase-documentation sub-genre, alongside the general-purpose (llm-wiki-gist) and heavyweight-personal (gbrain) instances.
A fourth shape arrived after this post (2026-08-31): the installable general-purpose one. Every
instance Chase names is either a hosted vendor product or one person's workflow. openkb
(pip install openkb, Apache-2.0) is neither — the pattern as a CLI anyone can point at their own
PDFs, with the wiki on their own disk. It also tests his scope boundary from the other side: he
excluded conversation state from wiki memory, and OpenKB's chat keeps sessions on disk beside the
wiki rather than compiling them into it, with /save as an explicit opt-in to promote a transcript to
a wiki page. The boundary is drawn where he drew it, by a builder who did not cite him.
Reflexive note
This hub (spokes.wiki) is itself a wiki-memory instance — an agent maintaining markdown raw/→wiki/
files as durable domain knowledge over curated sources. See synthesis for the "auto-maintaining wiki"
thread this corroborates.
Related
llm-wiki · harrison-chase · deepwiki · autowiki · openkb · retrieval-augmented-generation · open-knowledge-format · gbrain · synthesis