The Problem With How We Search Today
Ask a senior litigator why they chose one line of argument over another and you'll get a twenty-minute answer drawn from a career of wins and losses. Ask the firm's knowledge base the same question and you'll find the brief they filed, but not the thinking behind it. The answer lives in the partner's head. A new architecture, Andrej Karpathy's LLM Wiki, may finally give legal teams a window into a lawyer's mind.
Most legal teams that have attempted knowledge management have landed on the same basic tooling; a document management system, a keyword search bar, and good intentions. An associate searching for “breach of fiduciary duty Canada” will retrieve documents containing those exact words and miss the partner memo that framed the identical concept as “director loyalty obligations under the BCE standard.” Keyword search is structurally blind to meaning. After a few failed searches, lawyers abandon the system and start drafting from scratch.
Retrieval-augmented generation improved this picture. By converting documents into semantic vector embeddings, numerical representations of meaning rather than words, a RAG system can understand that a query about fiduciary duties and a memo about director loyalty are conceptually the same. That is a genuine leap forward. But RAG has its own structural flaw, and it is a serious one, it is perpetually forgetful (no persistent memory). Every time you query a RAG system, it retrieves relevant text chunks, synthesizes an answer, and then forgets the entire interaction. Ask a similar question next month and the system starts from zero. It searches for knowledge effectively, but it never builds knowledge.
Karpathy’s LLM Wiki: Knowledge That Compounds
Andrej Karpathy is a Slovak-Canadian AI researcher who co-founded and formerly worked at OpenAI. His LLM Wiki architecture changes the equation. Rather than treating AI as a search engine that passively scans a database of raw files, Karpathy’s model treats the AI as an active research librarian, one that reads, organizes, cross-references, and continuously maintains a living body of knowledge.
The architecture separates the system into three layers. The first is the raw source layer: original documents, transcripts, memos, and precedents, stored as immutable records. The second is the wiki itself, a curated, structured directory of plain-text Markdown files that the AI writes and maintains. The third is the schema: a set of standing instructions that governs how the AI organizes information, tracks provenance, flags contradictions, and maintains quality.
When a new document enters the system, say, a transcript of a senior partner’s lecture on recent appellate developments, the AI does not simply file it away for future retrieval. It compiles the information. It reads the transcript, extracts the legal principles, identifies the cases discussed, and integrates those insights into the existing wiki. It updates the relevant topic pages, creates new entity entries where needed, and inserts cross-references to related matters the firm has handled. The knowledge base doesn’t just grow larger. It grows smarter. Every ingested document makes every future query more precise. This is the critical distinction. A RAG system gives you search. The LLM Wiki gives you compounding institutional intelligence.
Capturing the Full Spectrum of Legal Judgment
What most knowledge management initiatives get wrong is that they focus exclusively on formal work product. The memos, the briefs, the contracts. But the most valuable knowledge a senior partner possesses is rarely captured in a filed document.
It is the article they read on a flight and mentioned in passing at a team meeting. The negotiation tactic they developed over fifteen years of dealing with a particular counter-party’s outside counsel. The notes they scribbled in the margin of a draft that explained why a specific clause was non-negotiable, not what the clause said, but the commercial reasoning and litigation risk behind their position. The CLE lecture where they walked junior associates through the strategic calculus of forum selection in cross-border disputes. The exposed blind spots, the quiet warnings about a particular jurisdiction’s procedural traps, the instinct that no formal training program has ever managed to teach. The same problem plays out inside a company's legal department. A general counsel's most valuable knowledge is often the escalation pattern they've worked out with a particular regulator, the reasoning behind why a particular vendor was approved despite a risky data processing term, or the internal politics of which business unit will push back on which position. When that GC leaves, their replacement inherits the contracts and the policies, but not the reasoning behind either.
The LLM Wiki architecture is designed to capture all of it. Automated transcription pipelines convert lectures and strategy meetings into raw text. Agentic workflows parse that text into structured legal frameworks, extracting issues, rules, analysis, and conclusions. The AI then compiles those structured outputs into the wiki, linking them to relevant practice areas, client histories, and precedent documents. A lawyer’s career worth of accumulated judgment becomes searchable, synthesizable, and permanent.
The Strategic Unlock: Matching Matters to Lawyers and Training the Next Generation
Once a legal team has built a genuine second brain, a range of new capabilities opens up. Matter intake, for example, shifts from an exercise in availability and seniority into genuine capability-matching. Instead of assigning work based on who has billed the most hours in a practice area, the firm can see who has actually handled the closest matters or built the deepest history with a specific regulator or opposing counsel.
The same system transforms how junior lawyers learn. An articling student or first-year associate no longer depends on catching a partner between meetings or stumbling across the right precedent by accident. They can query the second brain and receive not just the answer, but the reasoning, the strategic logic, the prior context, the specific considerations a senior partner would have applied. Decades of mentorship that previously moved through hallway conversations and red-lined drafts becomes available on demand. Training accelerates, quality floors rise, and the firm’s collective judgment is no longer bottlenecked by whoever happens to be in the office.