TL;DR — Chat context is a goldfish. An agent-written Obsidian vault is a personal wiki. Teams still need a durable knowledge base Software layer: linked pages humans can trust, permissions, search, and backups that restore files — not just another chat paste ritual.
If you hang around X right now, three threads keep colliding:
Karpathy's LLM Wiki pattern. Agents don't just dump notes — they maintain linked markdown pages that compound: immutable sources in, wiki pages out, claims that get revised instead of piled up.
Claude + Obsidian "second brain" repos. Free GitHub projects (
claude-obsidian,obsidian-skills, and friends) promising Claude Code stops forgetting everything the moment you close the terminal.The goldfish hangover. Every new chat wakes up empty. You re-explain who you are, what the product is, and which decisions already happened — then wonder why the agent invents a fourth architecture.
None of that discourse is fake. Linked agent-writable pages are a real upgrade over sticky notes in a chat window. Obsidian vaults are a decent substrate for personal agents.
But a pile of markdown files is not yet a team knowledge base — and "second brain" marketing often skips the part where humans still have to find, trust, and co-edit the same pages tomorrow.
Three memory models: ephemeral chat → connected wiki → shared knowledge space.
The three memory models people confuse
| Model | What it is | Wins when… | Breaks when… |
|---|---|---|---|
| Chat context | Paste a brief → answer → close the tab | One-off questions | Projects span days, people, or decisions you must not re-litigate |
| Agent LLM Wiki / Obsidian | Inbox → agent files + links → pages compound | One power user in a local vault | You need permissions, co-editing, whiteboards, or IT backups |
| Persistent AI knowledge base | Shared workspace: docs, links, DBs, optional AI on your pages | Knowledge must compound for the team | You only need a personal experiment |
That last row is the bar buyers should hold any "second brain" product to:
Does knowledge compound for the team — or only for the person who ran the agent last night?
A persistent AI knowledge base Software workspace is the system of record: not a chat log, and not only a personal vault.
Why the X discourse keeps missing the product
Compiling notes ≠ operating a wiki
"Stop storing notes, start compiling them" is a sharp slogan. Compilation still needs structure:
- Page history
- Search humans trust
- Permissions that keep draft strategy out of the all-hands space
- Backups that restore images as well as text
A self-hosted wiki comparison is less glamorous than a repo screenshot — and more predictive of whether the wiki survives quarter two.
Personal vaults don't automatically become team memory
Claude + Obsidian repos solve agent amnesia for one operator. They do not automatically solve:
- SSO and seat licenses
- Conflict-free co-editing
- "Who owns this page when the contractor leaves?"
Teams evaluating an Open source Notion style workspace are usually shopping for the latter — even when the Twitter thread that sent them there was about the former.
Reloading context is a product smell
If your daily ritual is pasting the same architecture doc into a fresh chat, you do not have a prompting problem.
You have a missing persistent layer.
The expensive part is not tokens — it is re-deciding things you already decided.
Buyer checklist: immutable sources, human-followable links, optional AI, clear license, full backup restore.
What a durable knowledge layer actually needs
Use this as a buyer checklist. Each item is something you can verify without believing a landing-page adjective.
1. Sources stay immutable; pages can change
LLM Wiki done right keeps raw inputs (transcripts, PDFs, export dumps) separate from the linked pages an agent maintains.
If every edit overwrites the only copy of a source, you cannot audit why a claim flipped.
2. Links humans can follow without the agent
Backlinks and collections only count if a new teammate can navigate them on day one.
If the graph only makes sense inside an agent's tool loop, you built a private cache — not a knowledge base.
3. Optional AI that reads the workspace (BYOK / your endpoint)
Chat UIs that upload a ZIP once are not the same as a workspace where AI is optional, keyed per workspace, and pointed at documents people already maintain.
For the stack view (model serving, chat, knowledge, agents), see self-hosted AI for teams.
4. An exit hatch that is not vibes
Licenses matter. Open core vs OSI open source vs source-available are different bets.
If "open" is part of your shortlist, read a license-checked open source Notion alternative comparison before you migrate a thousand pages on a badge alone.
5. Ops reality: backup = database + files + config
Self-hosting moves recovery to you. A database dump without uploads is a partial restore wearing a complete backup's clothes.
Pilot two tools for a week with real (non-sensitive) docs before you burn the Notion export.
Where AFFiNE fits (and where it doesn't)
AFFiNE is an open-source, local-first workspace that combines docs, an edgeless whiteboard, and database views — closer to "Notion + Miro in one place" than to a classic MediaWiki.
You can:
- Run cloud or self-host with Docker Compose
- Keep AI optional (BYOK or an admin-approved OpenAI-compatible endpoint)
- Let agents read workspace docs through a scoped MCP credential
Strong fit when:
- Your "LLM Wiki" pages need to live next to whiteboards, meeting notes, and team databases
- You want the persistent layer to be a real product — not a folder you hope the agent keeps tidy
Weak fit when:
- You only need a read-mostly handbook (BookStack is simpler)
- You require SAML / SCIM today
- You insist every line of the server is OSI-licensed MIT (AFFiNE is open core: MIT client, EE backend — read both license files)
A practical migration path from goldfish chats
Pick one project that currently lives in chat paste. Export the decisions into linked pages (problem, constraints, chosen approach, open questions).
Separate inbox vs wiki. Raw sources stay raw. Agent-maintained pages link out; they do not replace the archive.
Add one human ritual. Weekly: someone skims new pages, rejects bad links, and pins the three pages agents must read first.
Only then wire agents. MCP / vault skills / Claude Code hooks are multipliers on a clean base. On a messy base they multiply fiction.
If you need a team workspace, evaluate knowledge base Software options with the checklist above — not with a star count from a repo roundup.
Bottom line
The interesting part of the LLM Wiki / second-brain wave is not "AI can write markdown."
It is persistent, linked memory that compounds across sessions — instead of reloading context every chat like a goldfish with a credit card.
Personal Obsidian + Claude setups are a great personal experiment. Teams still need a durable AI knowledge base Software layer with permissions, search, and restore tests.
If your shortlist includes open tools that feel like Notion without the lock-in story, start from the homepage and read the license before you migrate: Open source Notion is a product category now, not just a Twitter punchline.
Build the wiki humans will still trust when the agent thread is gone. Everything else is autocomplete with amnesia.
Top comments (0)