Stop introducing yourself
to your own AI.
Cortex is one memory your agents share. Claude, Cursor, Claude Code, VS Code — each connects over the Model Context Protocol, with your consent, and starts already knowing you. No re-explaining. No pasting yesterday into today.
The Vision
A world where every agent you use — Claude, Cursor, ChatGPT, Claude Code — shares the same memory, because you gave it consent, not because one vendor owns your context.
BYOK, Zero Phone-Home
Self-host runs as a local stdio server — a single SQLite file, your own Gemini key, nothing phones home. The hosted server uses Personal Access Tokens so you choose which agents connect, and revoke access anytime.
Why Cortex?
You spend twenty minutes bringing an agent up to speed. It finally writes something good. Then the window closes, and all of it goes — the constraints, the preferences, the thing you already said twice.
Cortex keeps that context outside the chat, in a memory you own. Any MCP client can read it and write to it, so what you told Claude on Monday is there when you open Cursor on Thursday. And it holds up under measurement.
It's yours, not a vendor's
One memory, and you decide which agents see it. Hosted access runs on personal access tokens you can revoke at any time; self-host runs on your own machine with your own Gemini key. It isn't tied to one company's chat app.
The only published $/question
$0.0396 a question on LongMemEval_S: measured mean tokens per question, priced at Google's published Gemini list prices. Not an estimate. No competitor $/question has ever been measured — by them or by us.
One file, no Docker, no database
Self-host is a single SQLite file and nothing phoning home. The hosted service runs Postgres + pgvector. Same mcpServers config block either way: Claude Code, Cursor, Claude Desktop, VS Code.
Then we measured it.
0.932 across all 500 LongMemEval_S questions, at $0.0396 each. On LoCoMo, 0.813 across 1,986 questions at $0.0034. That's raw #2 of seven on one board and raw #3 of six on the other — below EMem, and the leaderboard is sorted so you can see it. On both, Cortex is the only system reading with a flash-class model, and the only one publishing a $/question at all.
All Cortex numbers are graded by an LLM judge = gemini-3.5-flash. The headline run used the same model family as both reader and judge, so a self-preference bias cannot be excluded; a cross-family re-grade is planned. The canonical LongMemEval judge is gpt-4o; a Gemini judge is likely more lenient, so a gpt-4o re-grade may lower these. Competitor rows use their own judges — the leaderboard is not apples-to-apples. Positioning is stated honestly: Cortex is raw #2 on LongMemEval_S and raw #3 on LoCoMo — the only memory system publishing $/question, and the only one at 0.9+ on LongMemEval_S with a cost-efficient-tier reader. Full benchmark breakdown
Six kinds of memory
Every memory your agents save is classified into one of six kinds — the taxonomy the dashboard and the memory brain both use.
Fact
Stable information about you or your world — identity, tools, background.
Preference
How you like an agent to behave — tone, format, style, do/don't rules of taste.
Instruction
A hard rule to always follow — 'never do X', 'always do Y'.
Project
Ongoing work, goals, and the things you're building.
Relationship
A person you know and how they relate to you.
Event
A dated thing that happened, anchored in time.
Underneath the six kinds, Cortex now organizes your memories into a connected, user-owned knowledge graph — entities and the relationships between them — that you can see and inspect, not just search one result at a time.
Open-core, by design
Apache-2.0 core
The memory engine, the MCP server, and the LongMemEval benchmark harness are all Apache-2.0 and public. The hosted control plane and this web dashboard are proprietary.
Your key, your data
Self-hosted, your Gemini key is read from the environment and used only to call Google's API — nothing else phones home. Local memory lives in a plain, unencrypted SQLite file at ~/.cortex/memory.db, so don't store secrets in it.