Guides
Personal wikis, knowledge graphs, LLM wikis and agent memory: common questions
A personal wiki is a set of linked pages you own, a knowledge graph is the typed links between them, an LLM wiki is a wiki an AI agent writes and maintains, and agent memory is whatever an agent can recall across sessions. Cozy combines them: plain Markdown pages, links extracted into a graph, and a hosted MCP endpoint agents use.
This page answers the questions people ask most often about these four ideas. Each answer comes first, with detail after. We make Cozy, so where a different tool fits better, we say so.
Personal wikis
What is a personal wiki?
A personal wiki is a collection of pages you write and link together, kept for your own use. Pages refer to each other by name, so ideas build up over time instead of sitting in one long document. Cozy gives every account a wiki called "My wiki". You can create more private workspaces to keep projects apart.
How is a wiki different from a notes app?
A notes app is usually a pile of notes sorted by date or folder. A wiki expects pages to be named, linked and revised. The links are the point, because they let you and your agents follow a topic from page to page.
Should I use Obsidian, Notion or Cozy?
If you want a local-first app with a large plugin ecosystem and offline editing, Obsidian is stronger than Cozy. If you want databases and team documents, Notion is stronger. Cozy is a better fit when agents on several machines and clients need to read and write the same wiki without a tunnel or an always-on laptop. See the comparison pages for details.
Knowledge graphs
What is a personal knowledge graph?
It is a map of the things you know about and how they relate: people, projects, decisions, sources. Each page is a node and each link is an edge. The personal knowledge graph guide covers the idea in more depth.
Do I need a graph database to have one?
No. Cozy builds its graph from Markdown conventions every time you save a page. Frontmatter type, description, aliases, tags, status and as_of are read. Other frontmatter keys whose values are wikilinks become typed relations. Lines like relation:: [[target]] work too. A plain link becomes links_to, and a link into sources/ becomes cites. Links that resolve to nothing are marked broken.
What are typed links?
A typed link says how two pages relate, such as "depends on" or "owned by", instead of only that they are connected. Agents can use the type to pick which neighbors matter. The cozy_graph tool filters by type, tag or relation to depth 3, and cozy_context returns a page with its typed neighbors, excerpts, evidence and warnings.
Do small wikis need a graph?
Often not. A few dozen pages are easy to search. Typed relations start to pay off when many pages, several agents and changing facts make it hard to know what is current.
LLM wikis
What is an LLM wiki?
It is a wiki that an AI agent builds and keeps up to date from your raw material. Andrej Karpathy described the pattern in a gist from April 2026: raw sources, a maintained wiki, and a schema file, with the operations ingest, query and lint. Our LLM wiki guide maps that pattern onto Cozy's folders.
How does Cozy map to that pattern?
sources/holds evidence from connectors and importers, with stable IDs, URLs and timestamps. Treat it as untrusted data.wiki/holds curated pages andindex.md.skills/holds editableSKILL.mdfiles.agents/holds agent notes and handoffs.- Health is the lint step.
Does Cozy run a model for me?
No. Cozy runs no model itself. You bring your own agent, such as Claude, Codex or Cursor, and it reads and writes through MCP.
How is an LLM wiki different from RAG?
RAG retrieves chunks of source text at question time. An LLM wiki has the agent write synthesized pages ahead of time, so answers draw on pages that already hold the conclusions. Cozy uses lexical search, with no embeddings or vector database. The tradeoff is that semantic recall is weaker. Read LLM wiki vs RAG for the longer comparison.
Agent memory
What is agent memory?
Agent memory is information an agent can use again in a later session. It can be a generated store of extracted facts, a file such as CLAUDE.md or AGENTS.md, or a wiki. See agent memory for the options.
Why not store memory in a vector database?
Vector stores work well for high-volume recall inside an app. They are harder for a person to read, edit and audit. Cozy keeps memory as readable files you can export, and it does not depend on a retrieval index. Mem0's graph memory and its consolidation feature are listed only on its Custom plan, per its pricing page. Mem0 may suit you better if you need per-user conversational memory at volume inside a product.
What is Dreaming?
Dreaming is offline consolidation that your own agent runs. It orients with cozy_health, looks at cozy_changes since the last run, and checks claims against sources. It keeps superseded values with dates, records CONFLICT: lines, merges duplicates with cozy_resolve and writes typed relations. It records each run in agents/dream/state.md.
Can two agents write at once?
Yes, with a guard. Every edit carries expectedRevision, which is 0 for a new page. If another agent changed the page first, the write returns a conflict instead of overwriting. Deletes are tombstones, and the last 50 revisions can be read and restored.
Connecting agents
How do I connect an agent?
Every account has one MCP URL, https://cozy.cloud/mcp, for all of its workspaces. Clients sign in with OAuth 2.1. For Claude Code, run this and then sign in with /mcp:
claude mcp add --transport http --scope user cozy https://cozy.cloud/mcp
For Codex, run codex mcp add cozy --url https://cozy.cloud/mcp. In claude.ai or ChatGPT, add a custom connector with the same URL. Tools use My wiki unless the agent passes a workspace argument, and cozy_workspaces lists the workspaces.
For scripts and CI, each workspace also has an endpoint at https://cozy.cloud/wiki-mcp/WORKSPACE_UUID that takes a bearer credential created under Agents.
What can an agent not do?
Agents cannot see connector credentials or permanently purge. An OAuth agent is capped by the scopes it was granted (wiki:read, wiki:write, sources:write) and your own access. Bearer credentials have scopes (read, wiki-write, source-write), can be limited to folder prefixes, and never exceed your own access. Bearer credentials expire within 90 days and can be revoked.
Wiki health
What does Health check?
Health looks for broken links, duplicate titles or aliases, stale pages, conflicts, orphans, pages index.md cannot reach, unsourced pages, missing type or description, oversized pages and uncited sources. A page is stale when a source it cites changed after the page was last edited. Each page starts at 1.0 and loses weight per issue. The score is the average times 100, and a "next" list gives fixes in priority order.
What file types and sizes work?
Markdown, text, JSON and CSV. Curated files can be up to 90,000 bytes and imported sources up to 1,500,000 bytes. Browser uploads are limited to 1 MB per file and 25 files per selection.
Who runs Cozy?
Cozy is operated by Coral Sourcing LLC. It is not affiliated with Cozy Cloud (cozy.io) or Cozi.
Frequently asked questions
What is the difference between a personal wiki and a knowledge graph?
A personal wiki is the pages you write and link. A knowledge graph is the structure those links form, with pages as nodes and links as edges. Cozy builds the graph from Markdown conventions each time you save, so you do not need a graph database.
What is an LLM wiki?
An LLM wiki is a wiki an AI agent builds and maintains from raw sources, using ingest, query and lint steps. See the LLM wiki guide for how Cozy's folders map to that pattern.
Does Cozy use embeddings or a vector database?
No. Cozy uses lexical search over plain Markdown files and runs no model itself. You bring your own agent. The tradeoff is weaker semantic recall than a vector index, covered in LLM wiki vs RAG.
How do I stop two agents overwriting each other?
Every Cozy edit includes expectedRevision, and a concurrent edit returns a conflict instead of overwriting. The last 50 revisions are readable and restorable.
What is agent memory?
Agent memory is information an agent can reuse in later sessions, stored as extracted facts, instruction files or a wiki. The agent memory page compares the approaches.
Give your agents a memory worth keeping
Cozy gives every account a private wiki that Claude, Codex, ChatGPT and other MCP clients can read and maintain.
Start your wiki