By the end of this guide, your AI assistant will remember what it learns and be able to search it later. You will connect Korium once, save a memory, and bring it back in a fresh chat.
What you need before you start
Three things. Get these ready and the rest is quick.
- A Korium account. Kyroco sets this up for you. You will get an invite email with a link. If you already signed in before, you are set.
- Claude Code on your computer. Claude Code is Anthropic's command-line assistant. It is the tool that will use Korium as its memory. Install it first if you have not.
- Your Korium web address. Kyroco gives you a link to your Korium, both the sign-in page and the connection address. Keep it handy.
Right now, Kyroco creates your Korium account and sends you an invite. There is no public self-serve signup yet. If you do not have an invite, talk to Kyroco to get set up.
Sign in to Korium
Korium's web workspace is called Salon. This is where you will later browse your memories and manage your team.
- Open the Korium web address Kyroco gave you.
- If this is your first time, open the link in your invite email and set your own password. The invite already ties you to the right organization.
- Sign in with your email and password, or with Google or Apple.
Once you are signed in, you have proven who you are. That is all you need for the next step.
Connect Korium to Claude Code
Now you tell Claude Code where Korium lives, so it can use Korium as a memory. You do this once with a single command. Open a terminal and run:
claude mcp add --transport http korium https://<your-korium-address>/mcpThis adds Korium to Claude Code as a connector named "korium". Replace the part in brackets with the address Kyroco gave you, and keep the /mcp at the end.
The first time Claude Code uses Korium, it opens your web browser and asks you to sign in to Korium, the same sign-in from step 1. This is how Korium confirms the connection is really yours. When the browser says "Code delivered", switch back to your terminal: Claude Code finishes the connection there.
Some accounts use a token instead of the browser sign-in. A token is a secret key that proves who you are. If you have one, add it to the command like this, on one line:
claude mcp add --transport http korium https://<your-korium-address>/mcp --header "Authorization: Bearer <your-token>"Keep the token private. Treat it like a password.
Check the connection
Confirm Claude Code and Korium are talking. In your terminal, run:
claude mcp listThis lists every tool Claude Code is connected to and whether each one is working.
Look for korium in the list with a connected or healthy status. If it shows a failure, double-check the web address from step 2 and that you finished the browser sign-in.
Save your first memory
Now the fun part. Start Claude Code as you normally would, and ask it to remember something. Your assistant uses Korium's capture tool to save it. For example, type:
Please remember in Korium: our launch date is March 3, and Dana owns the rollout plan.Your assistant saves that to your Korium memory. The save happens right away, so you are never left waiting. Behind the scenes, Korium also starts sorting out the facts and how they connect, but you do not have to think about that.
Save real things you would hate to repeat: decisions, owners, dates, and why a choice was made. The more Korium holds, the more useful its answers become.
Naming your agents and your work
Every memory records two extra things: which agent saved it and what kind of work it was. That is what lets you ask later "what did my reviewer decide?" instead of just searching everything.
You choose these names. There is no fixed list to pick from, and Korium will not reject a name it has not seen before. Type whatever matches how you actually work.
Suggested names to start with
If you are not sure what to use, these are a reasonable starting point. They are suggestions only, so take the ones that fit and ignore the rest.
Agents: orchestrator, interviewer, classifier, researcher, architect, builder, tester, reviewer, investigator, packager, teacher
Kinds of work: classify, interview, research, design, build, test, review, package, teach, handoff
Names like reviewer keep working when your team or your tools change. A name tied to one specific assistant has to be renamed every time you swap it out, and old memories then no longer match.
Adding your own
Just use it. The first time an agent saves a memory under a new name, that name starts appearing in your list. Nothing to register, no setup step.
Please remember in Korium, as my "analyst" agent doing "forecasting": Q3 renewals are tracking 12% ahead of plan.Seeing what you have used
Ask your assistant to list them:
Show me the agent names in my Korium memory.You will get every name you have used, with how many memories carry it. This is how you spot a typo: if reviewer has 400 memories and reveiwer has 1, the second one is clearly a slip.
Fixing a typo or a rename
You do not edit your memories to fix a name. You tell Korium what to display instead:
In Korium, show "reveiwer" as "reviewer".From then on those memories read as reviewer. The same works when something gets renamed: point the old name at the new one and everything lines up.
Korium keeps exactly what was saved at the time and only changes what you see. That means the record of what actually happened stays honest, and you can undo any of it by removing the mapping. Only an admin on your workspace can set these, and every change records who made it.
Filing memories under a scope
Besides who saved a memory, you can file it under a scope: a dotted path that works like a folder, so later you can ask for everything under one project in a single question. For example, billing.meters and billing.invoices both sit under billing.
A scope is a path, not a sentence, and its format is strict: each part may only use the plain ASCII letters a to z (upper or lower case), digits, underscore, or hyphen, with dots between the parts. No spaces, no slashes, and no accented characters. So deploiement.hydra works, but déploiement hydra is refused.
Please remember in Korium under scope "launch.rollout": Dana owns the rollout plan.The memory itself can be written in any language, accents and all. The ASCII rule applies only to the scope path, the same way a web address is restricted while the page it points to is not. If a scope breaks the format, nothing is saved and your assistant gets an error stating the rule, so it can correct the path and try again.
Search your memory
This is the moment Korium earns its place. Close your current chat and start a fresh one, so the assistant has no memory of what you just said. Then ask:
Search Korium: who owns the rollout, and when do we launch?Your assistant uses Korium's search tool. Korium does not hand back a list of scraps. It writes a short, clear answer, in this case naming Dana and March 3, pulled from what you saved in a different session. That is recall working across chats.
Take a look at Salon
Everything you save is also visible in the Salon web workspace. Sign in again from step 1 and explore:
- Console → Browse shows your saved memories. Click one to see its details and the facts Korium pulled out of it.
- Console → Observability shows the health of your memory: how much you have saved and how fast searches run.
- Organizations lets you switch between the spaces you belong to, if you are in more than one.
Salon is also where operators manage teams, invites, and, for Kyroco staff, the sealed client vaults. You only see the parts your role allows.
What to try next
You have the basics. Here are three good next moves.
Index your codebase
Ask your assistant to "import code index" so it can search your code by meaning, not by guessing file names. This is a new Korium feature.
Explore the knowledge graph
Save a few connected facts, then open Console → Browse to see how Korium links them into a web of decisions and causes.
Read the full user guide
The user guide explains every feature in plain words, including what shipped this week.
Add memory to a real project
Use Korium in a project you are actually working on. Save decisions as you make them, and let your assistant recall them next session.