A few days ago, Karpathy dropped LLM Wiki. The AI internet exploded.
"RAG is dead."
"A true second brain for AI."
"Finally, persistent knowledge that doesn't rot."
Five thousand stars. Millions of impressions. Universal praise.
I need to say something that's going to piss people off.
Karpathy got the direction right. The implementation is a dead end. This isn't a hot take from someone who read the README. I've spent the last year building PVM (Persistent Vector Memory) — an alternative AI memory architecture from scratch. I know exactly why LLM Wiki will fail, because I tried that approach first and watched it die.
First, credit where it's due
Karpathy nailed the problem. RAG is broken in a way most people feel but can't articulate:
- You retrieve context → inject it into a prompt → the LLM responds → everything gets thrown away.
- No accumulation. No iteration. No personal knowledge that grows over time.
- Every session starts from zero.
His core insight — knowledge shouldn't be temporarily retrieved; it should be continuously compiled, structured, and self-evolving — is 100% correct. He's asking the right question.
The answer he gave is just wrong.
LLM Wiki is symbolic AI wearing an LLM trench coat
Strip away the marketing, and LLM Wiki does this:
- LLM reads your documents
- LLM extracts "facts" and "concepts"
- LLM structures them into a wiki
- LLM links them together
- LLM updates links as new information arrives
This is not new. This is symbolic AI with a better parser.
Ten years ago, we called this knowledge graphs. Expert systems. Cyc. The entire industry burned billions of dollars on this approach and collectively agreed it doesn't scale. The problem was never who structures the knowledge — human or machine. The problem is the structure itself.
Two fatal flaws that haven't changed
Flaw 1: Information loss at extraction time
Every time an LLM "compiles" your raw text into a structured wiki entry, it makes a decision about what matters and what doesn't. It drops the exact phrasing. It drops the context. It drops the hesitation, the hedging, the "actually, on second thought."
In a personal memory system, what gets dropped is often what you need most — the detail that seemed irrelevant six months ago but is now the key to a decision.
LLM Wiki doesn't store your knowledge. It stores an LLM's summary of your knowledge. Those are different things.
Flaw 2: Errors compound, permanently
RAG makes a mistake → wrong answer this session → gone next session.
LLM Wiki makes a mistake during extraction → wrong fact enters the wiki → gets linked to other facts → gets used as context for future extractions → the error replicates itself through the graph.
This is the classic symbolic AI death spiral. Single-point error, global contamination. Ten years ago it was humans making the errors. Now it's LLMs. The failure mode is identical.
Karpathy himself acknowledges the scale ceiling: "Tested on ~100 documents in a single domain." The moment you cross into real personal knowledge — multi-domain, messy, contradictory, accumulated over years — the symbol graph explodes.
LLM Wiki vs PVM: same goal, opposite philosophy
The question both systems ask: how do we give AI persistent, evolving memory?
LLM Wiki's answer: Let AI understand the world for you, then store its understanding.
PVM's answer: Don't let AI understand anything. Store the evidence. Let the system adapt.
| LLM Wiki | PVM | |
|---|---|---|
| What gets stored | LLM-compiled summaries | Original text, untouched |
| Information loss | Inevitable — extraction is compression | Zero — append-only, nothing discarded |
| Error handling | Errors are permanent, spread through links | Dynamic weight decay, self-correcting |
| Scale ceiling | ~100 documents, single domain | Hardware-limited only |
| Dependency | Cloud LLM required for compilation | CPU-only, offline, zero dependencies |
| Philosophy | AI understands for you | System remembers for you |
The car analogy
LLM Wiki is a beautifully engineered 1.2L manual transmission sedan. Gorgeous to look at. Immaculate structure. Every gear shift requires conscious effort. Every boundary must be manually enforced. Push it past 100 km/h and the engine sputters.
PVM is a V12 automatic. It doesn't look as pretty in a diagram. There's no elegant tree structure. But it handles arbitrary load without complaint, self-corrects without human intervention, and doesn't ask you to maintain schema discipline at 2 AM.
The engineers who love LLM Wiki love it because it scratches an aesthetic itch — the fantasy of perfectly ordered knowledge. I get it. I have that itch too. But your life's knowledge is not a wiki. It's messy. It contradicts. It evolves. A system that forces it into a clean tree will break.
The real question
What do you actually want from an AI second brain?
An LLM that summarizes your life into a tidy, structured, aesthetically pleasing fiction?
Or a system that preserves everything you've ever thought, read, and decided — raw, untampered, permanently traceable — and gets better at surfacing the right parts the more you use it?
If you want the first one: LLM Wiki is beautiful. Use it for small-scale note organization. It's great at that.
If you want the second one: PVM exists. It's open source. It runs on your machine. It doesn't guess what matters — it lets time and your actual behavior decide.
PVM: https://github.com/rocky007cn/vector-as-memory
MCP server (works with Claude/Cursor): https://github.com/rocky007cn/ace-memory
Don't take my word for it. Run the code. The architecture doesn't lie.
Top comments (0)