Cortex documentation
Portable, agent-agnostic memory over MCP. Install it, try it live, and call it — everything here is public to read; running things needs a key.
A real run, captured and replayed: Google Gemini and ChatGPT write and recall the same memory through one key, in both directions. See more demos →
What Cortex actually is
Four properties define what Cortex gives you today.
Memory lives with you, not the model. Carry it from one agent to the next instead of re-teaching every tool from scratch. One personal key reads and writes the same store from curl, the Gemini CLI, a ChatGPT Action or Claude Code.
Relevant memory is pulled in and written back for you. The agent remembers without being told to, so there is no "use memory" step every turn.
An entity-relationship graph sits over hybrid keyword and semantic retrieval, so Cortex answers the indirect questions a flat document search cannot.
Cortex sits beside execution, never inside it. You can read and audit your memory, and it never quietly steers the agent. The labs won't offer this, because portable memory kills model stickiness, so Fifth9 ships it as a side layer, out of the execution flow.
One memory, every client
Claude, Google Gemini, ChatGPT and curl ring one Cortex store. A value is written once, recalled from every client, then forgotten for all — and at one point two clients act on the same memory at the same time. Each turn plays in three beats: the prompt, the memory fetch, then the answer.
Auto-playing. Press B/E for the business or engineer view, R to replay. More demos → · API reference →