# AI memory and personal knowledge glossary

> Plain definitions of 30 terms used in AI agent memory and personal knowledge management, from LLM wiki and MCP to RAG, provenance and memory consolidation.

Updated 2026-11-09. Canonical: https://cozy.cloud/glossary

> This glossary defines about 30 terms used when people give AI agents durable memory and keep personal knowledge in linked notes. Each entry starts with a one or two sentence definition, then notes how the idea shows up in Cozy where it applies. Terms are listed from core concepts to tooling and maintenance.

We make Cozy, a private wiki and cloud file space for each person and their AI agents. Where a term maps onto how Cozy works, we say so. Where it does not, we say that too. For a longer treatment of the central idea, see [LLM wiki](/llm-wiki).

## Core concepts

### Personal knowledge graph
A personal knowledge graph is a set of your notes and facts connected by explicit links, so that pages are nodes and relationships are edges. In Cozy the graph is built from Markdown conventions on every save, with no separate database to maintain. See [personal knowledge graph](/personal-knowledge-graph).

### LLM wiki
An LLM wiki is a folder of linked Markdown pages that an AI agent reads, writes and maintains on your behalf. It keeps raw sources, curated pages and agent instructions separate. Cozy uses `sources/`, `wiki/`, `skills/` and `agents/` folders for this.

### Second brain
A second brain is a personal system for capturing and organizing information outside your head. The term is popular in note-taking circles and usually refers to a human-maintained set of notes.

### PKM
PKM stands for personal knowledge management: the habits and tools you use to collect, connect and retrieve what you know. Tools range from paper notebooks to linked-note apps.

### Digital garden
A digital garden is a collection of notes kept public or semi-public and tended over time instead of published once. Pages are allowed to be unfinished and grow as links accumulate.

### Wiki
A wiki is a set of pages that link to each other and can be edited by many contributors, with a history of changes. Cozy gives every account a wiki called "My wiki" and you can create more private workspaces.

### Note
A note is a single unit of captured information, usually one page or file. In Cozy a note is a Markdown, text, JSON or CSV file.

## Links and structure

### Wikilink
A wikilink is a link written in double square brackets around a page name. Cozy resolves name links by path, file stem, title or alias, ignoring case and accents. Unresolved ones are reported as broken.

### Backlink
A backlink is the reverse of a link: a list of pages that point to the page you are viewing. Cozy exposes this to agents through the `cozy_backlinks` tool.

### Typed relation
A typed relation is a link that carries a label, such as "supersedes" or "owned by", instead of just saying two pages are connected. In Cozy, a frontmatter key whose value is a wikilink, or a line like `relation:: [[target]]`, becomes a typed relation. See [typed links](/typed-links).

### Frontmatter
Frontmatter is a block of key-value metadata at the top of a Markdown file. Cozy reads type, description, aliases, tags, status and as_of from it.

### Alias
An alias is an alternative name for a page, so links using either name reach the same page. Cozy also flags duplicate titles and aliases in its health check.

### Orphan
An orphan is a page that nothing links to. Cozy lists orphans, and pages that `index.md` cannot reach, as health issues.

## Agents and protocols

### MCP
MCP is the Model Context Protocol, an open standard that lets AI clients call tools and read resources on a server. Cozy hosts one MCP endpoint per account at `https://cozy.cloud/mcp`, over stateless Streamable HTTP, that reaches all your workspaces. Each workspace also has its own endpoint for scripts that use a bearer credential. See [MCP server](/mcp-server).

### Agent memory
Agent memory is information an AI agent can keep and reuse across sessions, instead of starting each conversation from nothing. Cozy stores it as files you can read and edit. See [agent memory](/agent-memory).

### Context engineering
Context engineering is the practice of deciding what information goes into a model's context window and in what form. Good memory systems return a page with its neighbors and evidence rather than whole archives. Cozy's `cozy_context` does this.

### Context window
The context window is the amount of text a model can consider at once. Anything outside it is invisible to the model unless it is retrieved and inserted.

### Tool call
A tool call is a request from a model to run a named function with arguments and get a result back. MCP defines how clients discover and invoke tools.

### Bearer credential
A bearer credential is a secret token that grants access to whoever presents it. Cozy's are created under Agents for one workspace, used with that workspace's MCP URL, and are random, shown once, stored as SHA-256 digests, revocable, and expire within 90 days. They suit scripts, CI and headless agents.

### OAuth with PKCE
PKCE is an extension to the OAuth authorization code flow that protects the code exchange. Cozy's account MCP server uses OAuth 2.1 with dynamic client registration, S256 PKCE and refresh tokens. You approve the agent on Cozy's consent page.

## Retrieval

### RAG
RAG, or retrieval-augmented generation, fetches relevant text at question time and gives it to a model to answer from. It is usually built on chunked documents and a vector index. Cozy does not use one. See [LLM wiki vs RAG](/llm-wiki-vs-rag).

### Embeddings
Embeddings are numeric vectors that represent text so that similar meanings sit close together. They power semantic search. Cozy uses no embeddings or vector database, and its search is lexical.

### Lexical search
Lexical search matches the words in a query against the words in documents. It is predictable and exact, but misses synonyms unless the page uses them. Cozy paginates results with a cursor.

### Vector database
A vector database stores embeddings and finds nearest neighbors quickly. It is a common component of RAG systems.

## Maintenance and trust

### Provenance
Provenance is the record of where a piece of information came from. In Cozy, links into `sources/` count as citations, and sources keep stable IDs, URLs and timestamps. See [source citations](/source-citations).

### Lint
Lint means automated checks that find problems in a body of content, the way a code linter finds bugs. Cozy's health report checks for broken links, stale pages, conflicts, orphans and missing metadata. See [wiki health](/wiki-health).

### Staleness
A page is stale when a source it cites changed after the page was last edited. Cozy flags this so a person or agent can re-verify the claim.

### Memory consolidation
Memory consolidation is an offline pass that merges duplicates, resolves conflicts and condenses what was learned. In Cozy this is called dreaming and is run by your own agent. See [memory consolidation](/memory-consolidation).

### Dreaming
Dreaming is Cozy's name for its consolidation routine. The agent checks health, reviews changes since the last run, verifies claims against sources, writes typed relations, and records the run in `agents/dream/state.md`.

### Revision
A revision is one saved version of a page. Cozy edits require an expected revision, concurrent edits return a conflict, and the last 50 revisions can be read and restored.

## Frequently asked questions

### What is the difference between an LLM wiki and RAG?

RAG retrieves chunks of text at question time, usually through embeddings. An LLM wiki keeps curated, linked pages that an agent maintains ahead of time. Cozy uses lexical search with no vector database. See [LLM wiki vs RAG](/llm-wiki-vs-rag).

### What does MCP stand for?

MCP stands for Model Context Protocol, an open standard for connecting AI clients to tools and data. Cozy hosts one MCP endpoint per account, at https://cozy.cloud/mcp, with optional per-workspace endpoints for bearer credentials. See [MCP server](/mcp-server).

### Does Cozy use embeddings?

No. Cozy uses no embeddings or vector database, and its search is lexical. The graph comes from links and frontmatter in your Markdown.

### What is memory consolidation?

It is an offline pass that merges duplicates, flags conflicts and keeps the memory tidy. In Cozy your own agent runs it, since Cozy runs no model itself. See [memory consolidation](/memory-consolidation).
