Cognitia AI

Verbraucher-KI-Speicherschicht

Dokumentation

The right memory for every agent

Cognitia is a user-authorized memory MCP server. Your client asks for origin groups, the owner approves Off / High-level / Detailed, and you work against their graph — inspectable, revocable, and not locked inside one chat product.

A permissioned read/write loop for user memory

Apps do not get blanket access. The user approves five origin groups and a duration. Tokens can be revoked.

1

Paste the MCP URL

Point the client at the Cognitia MCP server. It discovers OAuth on its own. Users do not paste a secret.

https://www.cognitia-ai.com/api/mcp

2

The user sets scope and a duration

They choose Off, High-level, or Detailed for five origin groups, and 7 days, 30 days, 90 days, or until they revoke.

scope=memory

3

You receive a token for that grant

Access lasts until the duration they chose. An expired grant stops reads and writes and is not deleted.

POST /api/mcp/oauth/token

4

Query or store through MCP

unified_memory_agent, unified_memory_retrieve, unified_memory_agent_config, unified_memory_store, and agent_actions_ledger run on the signed-in graph.

POST /api/mcp/jsonrpc

5

They can revoke anytime

Connected apps and origin grants stay under user control. Tokens can be cut without rebuilding your product.

Made for when context matters

Claude, Cursor, ChatGPT

Bring the user’s graph to the model they already use.

1

They connect Cognitia as an MCP server in the client they prefer.

2

Origin grants decide whether that client sees knowledge-graph memory, raw artifacts, or neither.

3

The same graph grounds chat, coding agents, and follow-ups without re-explaining their life.

Your product, with their memory

User-authorized memory for your agent or app — not a copy of their inbox on your servers.

1

Send them to Connect Cognitia, or have a headless agent show a device code.

2

Call MCP tools with the token you received for that user.

3

Let them review and revoke in connected apps when they are done.

Which tool to call

Personal initialize returns this router as instructions, plus memory.ask / memory.retrieve / memory.remember prompts and cognitia://status + cognitia://profile.

Call action=list once per session. See origin grants, last_ingest_at, and local_files.

People, facts, or a profile. Fast, no LLM. Iterate a node with entity_id.

A written answer, today's inbox, or calendar. Set include_evidence when you will cite.

Only a durable note the user asked to save. Keep source_id. Cannot delete.

Record or query what a connected agent did. mode is record or query.

Featured clients — copy JSON, then OAuth

Claude, Cursor, ChatGPT, Copilot, Gemini, Windsurf, Zed, and Cline. Install Cognitia as a memory MCP server from MCP & backups after sign-in.

Use Cognitia in

Install Cognitia as a memory MCP server in the client you already use. One JSON or URL, then OAuth. Scripts or clients without OAuth can use a personal API key below.

Claude

Desktop, claude.ai, and Claude Code. Highest-conversion memory client.

Cursor

Same graph in the editor. One JSON block; OAuth on first tool call.

ChatGPT

Apps / Developer Mode MCP connector. Same user-authorized memory endpoint.

GitHub Copilot

VS Code / Copilot agent MCP. Default editor for many power users.

Gemini

Gemini CLI and Google AI Studio MCP. Same user-authorized memory endpoint.

Windsurf

Cascade MCP. Cursor-class editor with the same memory graph.

Zed

Native remote MCP in settings.json. Omit Authorization so Zed runs OAuth.

Cline

VS Code / Cursor agent. MCP JSON in Cline settings, then OAuth.

Public MCP endpoint

Clients connect over JSON-RPC with OAuth. Origin grants are Off / High-level / Detailed (API: off / kg / raw) on Email, Finance, Social, Storage, and MCP & Companion. Ask searches those groups through an inner agent (graph, Pinecone, official live) rather than ungated “search everything.”

https://www.cognitia-ai.com/api/mcp

Listed as a memory MCP server on mcpservers.org. OAuth lives at /api/mcp/oauth/*.

Connect button

Hosted apps use a link. The client id is an HTTPS metadata document on your own host. PKCE is required.

https://www.cognitia-ai.com/api/mcp/oauth/authorize
  ?response_type=code
  &client_id=https://your.host/.well-known/cognitia-client.json
  &redirect_uri=https://your.host/oauth/callback
  &code_challenge=<S256>
  &code_challenge_method=S256
  &scope=memory

Device pairing

A headless agent asks for a device code. The user enters it at this page and approves the same grant. /device

curl -X POST https://www.cognitia-ai.com/api/mcp/oauth/device \
  -H "Content-Type: application/json" \
  -d '{"client_id":"https://your.host/.well-known/cognitia-client.json","scope":"memory"}'

Client metadata document

{
  "client_id": "https://your.host/.well-known/cognitia-client.json",
  "client_name": "Errand",
  "redirect_uris": ["https://your.host/oauth/callback"],
  "grant_types": ["authorization_code", "refresh_token"],
  "token_endpoint_auth_method": "none"
}

Registration stays available for clients that cannot host a metadata document.

curl -X POST https://www.cognitia-ai.com/api/mcp/oauth/register \
  -H "Content-Type: application/json" \
  -d '{
    "client_name": "My MCP Client",
    "redirect_uris": ["https://myapp.com/oauth/callback"],
    "grant_types": ["authorization_code", "refresh_token"],
    "token_endpoint_auth_method": "none"
  }'

Why this exists now

AI products get better the more they know about someone, so every assistant is building a private profile. That profile does not travel. Cognitia is the graph the person owns; MCP is how Claude, Cursor, or your app asks for a slice with a purpose and an expiry they can revoke.