CortexMind CortexMind is one long-term memory for your AI tools, on your own computer. Every conversation opens with what you decided, what you prefer and where you left off, in every tool you connect.
14-day free trial on the monthly plan. macOS, Windows and Linux.
CortexMind is a separate memory on your own computer that your AI tools connect to. These are the five places you would notice it, and the first three apply even if you only ever use one AI tool.
Your AI is told to save every exchange in full, so the decision you mentioned in passing and the error you fixed on Tuesday are still there months later.
You would notice it when: You know you discussed this, but nobody wrote it down.
What you decide in one project is there when you start the next, without copying notes between folders.
You would notice it when: You solved this in another repo last month.
A dashboard on your computer shows every saved memory. Change what matters, retire what went stale.
You would notice it when: You wonder why your AI just said that.
Say "remember that I want PR descriptions under 100 words" and it is there at the start of every conversation, in every connected tool.
You would notice it when: You have asked for the same thing in three tools this month.
Claude Code, Codex, Cursor, ChatGPT and claude.ai connect to one memory. What one of them learns, the others know.
You would notice it when: You start something in Claude Code and pick it up in ChatGPT.
First ask your AI what it remembers about you and where it keeps that. Then, in your first week, start in a different project and ask about something you decided in another one. Then ask the same question from a second tool.
Use ChatGPT or Claude in a browser or on your phone? Those apps connect from their own servers, so turn on the hosted tunnel in CortexMind. The computer running it has to be on at the time.
You explain the project again. You restate the rule you set last week. The agent works out a bug it already fixed. If that happens to you, even some of the time, this is the gap CortexMind covers.
It is a memory that outlives the conversation: your AI reads it before it plans and writes to it before it stops.
Read how it worksInstall once. Your agents boot with memory from the next conversation on.
Download the app for macOS, Windows or Linux and paste the license key from your receipt. The app lives in your tray and keeps your memory running in the background.
Connect the AI tools you use. Tools that run on your computer, like Claude Code, Codex, Cursor or Grok Build, connect directly with one config snippet from the app. The chat in claude.ai, Claude Desktop or ChatGPT reaches it through the secure link included with your license.
The agent boots with your standing rules and the last turns, searches memory as the topic moves, and saves the turn before it ends. You stop repeating yourself.
One prompt pasted into Claude Code, Codex, Cursor or any agent with a shell installs the app, connects the agent, moves the memories it already keeps into CortexMind, and interviews you to seed the rest.
Let your AI install itIt searches by keyword and by meaning at the same time. A second model then reads the best candidates against your question and ranks the closest matches first. The importance you set and the age of a memory settle the ties. Search stays quick as your memory grows: about a second on a desktop graphics card.
No language model sits in the write path. What you save is what gets stored, in full, on your disk.
Full-text search (FTS5 and BM25) and dense vector search run side by side and are fused with reciprocal rank fusion. A cross-encoder (bge-reranker-v2-m3) then reranks up to 60 candidates. Each result comes back with a relevance score, a type and a date.
memory_search("what to leave out of release announcement messages to users")
Customer-facing release announcements name the benefit only, never the problem that was fixed
Public repo hygiene: never embed internal colleague names in GitHub release titles or public artifacts
The boot call returns your standing rules, the preferences and conventions you promoted to the top, plus where the last conversation left off. The agent has them before you type a word.
When the card grows too long, CortexMind tells the agent to ask you to prune it. Your memory stays sharp instead of piling up.
memory_init
Standing rules
Where we left off
Migration PR 412 merged, cutover window still to schedule.
Sarah asked for the science lesson as a PDF and a phone page.
With a memory that keeps growing, the model matters less and what it knows about you matters more. Every conversation leaves it a little sharper, and the improvement compounds.
Say "remember that I want PR descriptions under 100 words" and it is a standing rule from the next conversation on. Setup adds one short block to your instructions file; after that you do not hand-write rules.
Switch models, switch tools, keep the memory. What makes your agent good is what it has accumulated about you and your work, and that stays yours.
Point it at a topic, a codebase or a pile of documents and let it read. What it saves is there for every agent, every conversation, from then on.
A private dashboard on your machine. Browse episodes, inspect facts, search the way your agents do, raise what matters and archive what went stale.
Only your own computer can open it. Nothing on it is sent anywhere.
On purpose: your memory lives in one file in your home folder, not on a server. No account, no cloud sync, no telemetry on your content. Back it up like any other file.
Apple Silicon runs the models on Metal. NVIDIA, AMD and Intel cards run them on Vulkan. No graphics card runs them on the CPU, slower and still correct. Nothing to configure.
The license includes a secure link called a hosted tunnel. Turn it on and claude.ai, ChatGPT or any hosted agent reaches your memory over HTTPS, while the data stays home.
A card is required for the trial and nothing is charged for 14 days. Cancel from your customer portal at any time.
14-day free trial
$10per month USD
Save 17%
$99per year USD
One license activates one machine. Tax is handled at checkout by Polar, the merchant of record. Your license key lives in your customer portal.
Already have a license? Pick your platform.
Released November 11, 2026
All releases on GitHubWithout a graphics card the models run on the CPU and everything still works. With one, CortexMind times your CPU against every GPU it can see on first launch and keeps the fastest.
No. The embedding and reranking models run on the CPU when nothing else is available. A GPU makes saving and recall faster on a large memory, and CortexMind picks it up on its own: Metal on Apple Silicon, Vulkan on NVIDIA, AMD and Intel cards.
In one database file in your home folder. You can back it up, copy it to another machine or delete it like any file. Nothing is synced to a server.
Any tool that speaks MCP. Tools that run on your computer connect directly, with nothing extra to turn on: Claude Code (and the Code tab in Claude Desktop), Codex in the terminal or in the ChatGPT desktop app, Cursor, Grok Build and other coding agents. The chat in claude.ai, Claude Desktop and ChatGPT connects from their own servers, so those need the hosted tunnel, which you turn on in the app.
It is told to. The tools carry their own instructions: load your memory at the start of a conversation, search as the topic moves, and save before the turn ends. A short block in your CLAUDE.md or system prompt turns that into a habit, and the docs page has the block to paste. Everything it saves appears in your dashboard, so you can see what it kept.
Yes, that is the point. One daemon on your machine serves every client you connect, so what one agent learns the next one knows.
The app keeps working until the end of the paid period, plus a 72-hour offline grace. Your database stays on disk untouched. Subscribe again and everything is back.
The license check talks to Polar. If you turn on the hosted tunnel, requests from hosted agents pass through the relay to your machine over HTTPS. Your memories are never uploaded anywhere.
The monthly plan starts with a 14-day trial so you can decide before paying anything. For anything else, write to [email protected].
Your AI moves it over first. The setup prompt has it read what it already remembers, from memory files on disk or a built-in memory feature, and save each durable fact into CortexMind. Only then is the built-in memory switched off, and the prompt never asks it to delete anything. It has to be off because otherwise your AI keeps saving new things in the old place, and you end up with two memories that disagree. In most apps your agent tells you where the switch is and asks you to turn it off.
Only while you want the memory available. For most people CortexMind runs on the computer they are working on and starts with the session, so there is nothing to think about. If you want claude.ai, ChatGPT or the phone apps to reach your memory at any hour, the machine running CortexMind has to be on and reachable at that moment: leave a home computer running, or install CortexMind on a small VPS in the cloud so it is always online.
Through the hosted tunnel. Turn it on in the app and CortexMind gets an HTTPS address that claude.ai, ChatGPT or any remote MCP client can add as a connector, signing in with OAuth. The tunnel is EtherTunnel, a separate product: your CortexMind license includes a free EtherTunnel license with one tunnel. Discover EtherTunnel
On purpose. Your memory holds your decisions, your clients and how you work, so it stays on your machine instead of on someone else's server. No CortexMind server stores your memories, so there is nothing of yours for us to lose or read. Search runs on your own hardware: about a second on a desktop graphics card, about seven seconds on an M4 MacBook Air, longer with no graphics card. The trade-off is that the computer running CortexMind has to be on when you use it.
Yes, with one memory. Install CortexMind on one computer and turn on the hosted tunnel in the app. On the other computer, add the tunnel address to your AI tools and sign in. The computer running CortexMind has to be on. A license covers one install and its one memory database. Two separate memories work against the point, but if you do want them, that is one license each.
