The state of AI memory today
Every day, millions of people pour their thinking into ChatGPT, Claude, Gemini, and Cursor. We ask hard questions. We get back genuinely useful answers — strategies, code, frameworks, insights that took experts decades to develop.
Then we close the tab.
That answer is gone. Not literally — it sits in some chat history we'll never search. But functionally gone. We can't find it. We can't reuse it. We can't build on it.
The next day, we ask similar questions. We get similar answers. We close the tab again.
This is happening at civilizational scale. Trillions of tokens of high-quality, AI-assisted thinking, evaporating into chat histories nobody returns to. The world's most expensive forgetting machine.
The industry's response so far is what I call extracted memory — services like Mem0 and Letta that watch your conversations and extract facts the AI thinks are important. "Sarah is vegetarian. Sarah lives in Seoul. Sarah is interested in LLM evaluation."
Mem0 and Letta are excellent at what they do. They solve a real problem. But they answer a different question than the one I want to answer.
They ask: what should the AI remember about you?
I want to ask: what do you want to remember?
These are not the same question. The first is about inference. The second is about authorship.
Why authorship matters
Memory is not just data. Memory is identity.
What you remember shapes who you become. What an organization remembers shapes what it can do. This was true in the age of paper, true in the age of databases, and is more true in the age of AI than ever before.
When you let an AI extract your memory, you let an AI define what mattered. You let an algorithm decide which thread of yesterday's thinking is worth carrying forward, which insight to compress into a fact, which piece of yourself to keep.
That's a strange thing to outsource.
Some people will outsource it gladly. The convenience is real. But for those of us who think carefully about what we want our future selves to know — for those of us who treat our knowledge as a craft, not a byproduct — there should be another option.
That option is memory.wiki.
What memory.wiki is today
If you visited memory.wiki right now, you'd see what looks like a markdown publishing tool — and it is.
You can capture markdown from anywhere: ChatGPT, Claude, Gemini (via Chrome extension), GitHub repos, Notion pages, Obsidian vaults, any web URL, your terminal (cat README.md | mw publish), VS Code, your Mac clipboard. You can edit it in a beautiful WYSIWYG editor — no syntax friction, no install required. You can share it with a permanent URL that anyone can read in the browser, that any AI can fetch as context.
It's a publishing tool. It works. People can use it today.
And it's grown past the publishing layer. Every saved doc folds into a personal hub that auto-publishes a wiki manifest (index.md, SCHEMA.md, log.md), maintains a concept index across your library, and exposes a recall API any AI can query. The role-split is the load-bearing idea: you set the direction, memory.wiki structures the URL, any AI reads it.
Shipped so far:
- A unified markdown render pipeline (markdown-it) shared across every surface
- A web editor with WYSIWYG
- A Chrome extension for any AI chat
- VS Code extension, Mac desktop app, CLI, MCP server
- Five import surfaces: GitHub, Notion, Obsidian, URL, files
- Hub manifests — index.md / SCHEMA.md / log.md / llms.txt per hub
- Hub recall API with optional Haiku-based reranker
- Concept index + Related-in-your-hub + Needs-review lint
- Auto-classified doc intent (note / definition / decision / etc.)
I shipped this fast because I had a clear primitive: the markdown URL. Every surface points to the same thing. Every surface composes with the others.
The bigger bet
The bigger bet is that markdown URLs are the right substrate for AI-era knowledge.
Not as a publishing tool. As infrastructure.
LLMs read and write markdown natively. It's the lingua franca they were trained on. When ChatGPT outputs structured information, it outputs markdown. When you paste context into Claude, you paste markdown. When agents communicate with each other, the natural format is markdown. This is not changing. It's compounding.
Humans also read markdown natively. Plain text formatted lightly is how we've taken notes for centuries. It's how we'll keep taking them. No proprietary format will displace it.
URLs are the simplest possible interface. Anyone can paste them. Any agent can fetch them. They cross every boundary — operating systems, applications, AIs, time zones, decades.
Andrej Karpathy put it the cleanest: "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." He drew it for the Obsidian shape. We drew it for the URL shape. Same insight, different deploy target — local files vs. a public address every AI can fetch.
If LLMs write markdown, humans read markdown, and URLs cross every boundary, then the natural primitive for AI-era knowledge is a markdown document at a URL.
What's coming next
The memory layer is now live. Hub URLs, recall API, concept index, and the wiki manifests all shipped. What's left is the part that turns memory.wiki from a tool people learn into one they reach for without thinking.
You should be able to take what you've authored and deploy it as context to any AI, anywhere — without remembering how memory.wiki works.
Deeper ingest: Bear, Apple Notes, Roam, Drive. Native ingestion from every place people already keep notes, with the same dedup contract that makes re-importing safe.
Sharper recall: latency budgets on the cross-encoder rerank, a hybrid scoring that the user can dial without reading docs, and citations that visibly improve the answer instead of just appearing under it.
Cleaner curation: today the Needs-review lint surfaces orphan docs and likely duplicates. Next: stale claims, contradiction detection, missing cross-references — the lint signals Karpathy gestured at for his wiki, applied to yours.
Plus the mundane work that ships a real product: payment, custom domains, branding, a public hub directory worth pointing other people at.