AI Meeting Assistant vs LLM Wiki: Which Do You Actually Need?
10 min read

An LLM wiki is a knowledge system where an AI agent reads your source documents once and compiles them into interlinked markdown pages it maintains over time, rather than re-retrieving raw fragments on every query. Andrej Karpathy described the pattern in a GitHub gist in April 2026. It cannot replace an AI meeting assistant, because a wiki has no way to capture a live conversation. The two are complementary: the assistant captures what was said, the wiki synthesizes it into durable knowledge.
A question worth answering directly: should an LLM wiki replace your meeting assistant? The short answer is no, and understanding why clarifies what each one is genuinely for.
This guide covers what the LLM wiki pattern actually is, where it came from, what it does well, why it cannot capture meetings on its own, and how teams combine both.
What Is an LLM Wiki?
An LLM wiki is a knowledge base an AI agent builds and maintains itself. The agent reads your source material once, synthesizes what it learned into structured markdown pages, and keeps those pages current as new sources arrive.
Andrej Karpathy, an OpenAI co-founder and former Director of AI at Tesla, described this pattern in a GitHub gist published in April 2026. The gist is short, and it has been widely forked and discussed since.
The core argument is about when the thinking happens. A retrieval system does its work at query time, searching raw documents whenever you ask something. An LLM wiki does the work at write time, compiling knowledge once so queries read a finished artifact instead of re-deriving an answer from scratch.
The Three-Layer Structure
The pattern separates three things that usually get mixed together.
Raw sources sit in their own layer and stay immutable. The agent reads them and never modifies them, so the original material remains intact.
The wiki is a directory of markdown pages the LLM owns entirely. It writes them, updates them, and links them to each other. Nothing in this layer is hand-written.
The schema is a configuration document telling the agent how to maintain the wiki: what conventions to follow, how pages should be structured, what workflows apply.
That separation is what makes the system durable. Sources stay trustworthy because nothing rewrites them, and the wiki stays coherent because one consistent set of rules governs how it grows.
How an LLM Wiki Differs From RAG
Factor | Retrieval (RAG) | LLM Wiki |
When synthesis happens | At query time, every time | Once at write time, then maintained |
What the model reads | Raw document fragments | Compiled, interlinked pages |
Infrastructure required | Vector database, embeddings, retrieval pipeline | A markdown folder and an agent |
Knowledge accumulation | Does not compound, each query starts fresh | Compounds, pages improve as sources arrive |
Main limitation | Retrieval quality varies per query | Breaks down once the wiki exceeds available context |
The scaling limit is the honest caveat. A wiki works well while it fits comfortably in a model’s context window. Past that point, the pattern needs its own retrieval layer, which reintroduces the problem it set out to avoid.
Why an LLM Wiki Cannot Replace a Meeting Assistant
This is the part the original question gets backward, and it comes down to one practical fact.
An LLM wiki has no microphone. It processes documents you give it. It cannot join a call, hear what people say, identify who said it, or capture a decision made out loud in a conversation that produces no document at all.
A meeting assistant exists precisely to solve that capture problem. It converts a live, spoken conversation into a record that did not previously exist in any written form.
Without capture, there is nothing for a wiki to compile. The meeting happened, people made decisions, and no source document was ever created. A wiki cannot synthesize a conversation nobody wrote down.
How the Two Actually Fit Together
The useful framing is capture and synthesis as two halves of one pipeline.
Your meeting assistant handles capture. It produces accurate transcripts and summaries from live conversations, turning spoken discussion into written sources.
An LLM wiki handles synthesis. It reads those transcripts alongside your other documents and compiles durable pages about projects, decisions, and recurring topics that span many meetings.
The output is different from either alone. A meeting assistant tells you what happened in Tuesday’s call. A wiki built on months of those transcripts tells you how a decision evolved across a quarter, which no individual summary contains.
Who Should Actually Build an LLM Wiki
Honesty matters here, because this pattern is not for everyone.
It suits individuals and small technical teams comfortable working in markdown files and running an agent with filesystem access. Karpathy’s own version reportedly grew to around a hundred articles and several hundred thousand words, none of which he wrote by hand.
It suits people whose knowledge genuinely compounds: researchers, engineers maintaining complex systems, anyone whose work benefits from accumulated context rather than one-off answers.
It does not suit a team that simply wants meeting notes. The setup effort is real, the maintenance is ongoing, and a good meeting assistant already covers that need without asking anyone to manage a markdown vault.
Enterprise Use Cases
A Team Wanting Knowledge That Compounds Across Meetings
Problem: A team has accurate individual meeting summaries but no way to see how a decision developed across many conversations over months.
Solution: Feeding meeting transcripts into an LLM wiki alongside other project documents, letting the agent compile pages that span the full history.
Outcome: The team gains a durable view of how decisions evolved, which no single meeting summary could ever contain.
Someone Considering a Wiki Instead of a Meeting Assistant
Problem: Someone reads about the LLM wiki pattern and wonders whether it removes the need for a meeting assistant entirely.
Solution: Recognizing that the wiki has no capture mechanism, and that without transcripts there is no source material about meetings for it to compile.
Outcome: The person keeps capture in place and adds synthesis on top, rather than removing the layer that generates the sources.
A Researcher Managing Accumulated Context
Problem: A researcher accumulates papers, notes, and meeting transcripts faster than they can maintain a coherent mental model of the whole body.
Solution: Running an LLM wiki over that material, letting the agent maintain interlinked pages per concept and entity.
Outcome: The accumulated material becomes navigable knowledge rather than a growing pile of files nobody has time to reread.
Decision Tree: Do You Need a Wiki, an Assistant, or Both?
HintMint Capture and Synthesis Framework v1.0
HintMint framework for deciding how capture and synthesis fit your situation.
Step 1: Identify whether your gap is capture or synthesis. Missing records of conversations is a capture gap. Having records nobody connects is a synthesis gap.
Step 2: Solve capture first if it is missing. Synthesis has nothing to work with until sources exist.
Step 3: Assess your genuine appetite for maintenance. An LLM wiki requires ongoing upkeep and technical comfort with files and agents.
Step 4: Feed capture output into synthesis if you build both. Meeting transcripts become source material the wiki compiles.
Outcome: Completing this sequence puts the right tool against the right gap, rather than adopting a pattern that cannot address the problem you actually have.
Final Take
An LLM wiki is a genuinely interesting approach to AI memory, built on the idea that knowledge should be compiled once and maintained rather than re-derived on every query. It is not a replacement for a meeting assistant, because it has no way to capture a live conversation in the first place.
The clearest recommendation: keep capture in place, and consider a wiki on top if your knowledge genuinely compounds and you have real appetite for maintaining one. If you simply want reliable meeting notes, a meeting assistant already covers that without the overhead.
About Author
Muhammad Aatif Bashir is the Founder & CEO of RTC LEAGUE, a deep-tech company delivering enterprise AI communication solutions. With a strong business and telecom leadership background, he drives the vision behind TelEcho and HintMint, enabling organizations to scale intelligent customer interactions and enhance decision-making in high-stakes, real-time environments.


