Your notes files remember. Team0 keeps up.
Most people give their agents a second brain made of markdown: a file about themselves, notes on people, projects and tasks, and an instruction to keep them updated. It works at first. Then plans change, the files drift, and every agent reads its own copy. Team0 builds one understanding from your work, keeps it true, and shares it with every agent you use.
What people build today.
Almost every setup is a handful of markdown files and an instruction to keep them current. It is a good idea, and plenty of people run it well.
- A file about you
- about-me.md or USER.md: your role, your company, how you like to work, what you are focused on right now.
- Files for your work
- A note for each person, projects and tasks, a log of decisions, meeting notes and a note for each day.
- An instruction to tend them
- A line in CLAUDE.md or AGENTS.md that says where the files are: read them first, update them at the end of every session.
Plain text is yours, readable and free. If you work with one agent on one project, a notes file may be all you need, and you should keep using it.
What running it actually takes.
Making the files is the easy part. Ready-made kits take care of some of this, like the file layout, the instructions and a nightly tidy-up. They still need installing and connecting to each agent, and every update is still the agent’s call.
- 01
Decide the files
Which ones, what goes in each, how they are named. Get it wrong and the same fact ends up in three notes, or in none.
- 02
Write the instructions
Read these before every turn, update them after. Then write it again for every agent in its own format: CLAUDE.md for Claude Code, AGENTS.md for Codex, OpenClaw’s own files. And keep them in step.
- 03
Give every agent the folder
On your laptop, in a GitHub repo, in a synced drive. Each agent needs its own way in. Hosted agents like ChatGPT, Town or monday.com often have none, and setting one up is a project for most people.
- 04
Trust the agent to do it
Whether it reads the files, what it decides is worth saving, where it writes it, what it shortens: each one is the model’s call, every turn. A skipped update or a guess saved as fact looks exactly like a good one.
- 05
Type in what happened elsewhere
The meeting that moved, the email you got on Sunday, the doc a colleague changed. If it didn’t happen in a chat, the files don’t know until someone tells them.
- 06
Follow every change through
When Dan leaves, his tasks, the projects he owned and next week’s meeting sit in other notes. Updating his own note is easy. Finding everything that depended on him is the hard part.
- 07
Keep it small enough to read
The files outgrow what an agent loads. Someone prunes, summarises and archives, and the detail that mattered can go with it.
- 08
Check it
Read the files now and then, catch the wrong line, merge two versions of the same fact, and recover when one session saved over another.
Where files start to slip.
Even when all of that is done well. It shows once real work starts changing what the files say, and a second or third agent starts reading them.
- Plans change. The files go stale.
- Friday becomes Thursday. One note gets updated, two still say Friday, and an agent acts on whichever line it reads first.
- Heard looks like decided
- Every line weighs the same. Something a customer floated once on a call reads exactly like something you agreed.
- Each agent keeps its own version
- Unless every agent is wired to the same folder, Claude Code follows CLAUDE.md, Codex follows AGENTS.md, and OpenClaw loads its own profile and memory files. Even with one shared folder, two sessions saving at once can quietly replace each other’s work.
The same change, two ways.
Jordan emails on Tuesday to move Friday’s pilot review. The calendar confirms Thursday 14:00. Earlier, on a call, Jordan mentioned preferring the Growth plan.
Without Team0
Your markdown files
projects/northpeak.md - Pilot review with Jordan, Friday 11:00 - Jordan wants the pilot on the Growth plan people/jordan.md - Runs the Northpeak pilot. Review on Friday. daily/tuesday.md - Jordan emailed: can we move the pilot review to Thursday?
Two notes that still say Friday, one that says Thursday, and a passing remark that reads like a decision. What an agent does depends on which line it reads first.
With Team0
One understanding
- Statement
- Pilot review with Jordan, Thursday 14:00
- Status
- Confirmed by your calendar
- Source
- Jordan’s email: “Can we move the pilot review to Thursday?” and the calendar change that followed
- Said
- Tuesday 09:40
- Learned
- Tuesday 09:41
- Replaced
- Friday 11:00 (kept in history)
Heard, not confirmed
Jordan would like the pilot on the Growth plan. Said once, on a call.
What Team0 does instead.
One understanding underneath every agent, built and kept current without you tending it.
- 01
Built from your work
Team0 reads the sources you connect and what your agents save back. Nobody has to write anything down, and nothing waits for the next chat to be learned.
- 02
Every agent reads the same understanding
Claude Code, Codex, OpenClaw and Hermes read it before every reply and save when a turn ends. ChatGPT, Claude, Town, monday.com and Base44 read it when you ask. Tell one something, and the others already know.
- 03
Confirmed, heard, replaced, unknown
Kept apart, and your agents are told which is which. Something heard never overwrites something confirmed, and it fades out if nothing ever confirms it.
- 04
Changes are kept, not overwritten
A newer statement replaces an older one. The older one stays in history with what changed it, and every statement records when it was said, when Team0 learned it and the date it is about.
- 05
Checked before it is kept
Every claim goes through the same steps: who it is about, the exact line it came from, what kind of statement it is. A question saves nothing.
- 06
It keeps itself true
Checks run as each statement is saved and again every night. Contradictions are caught and put to you through your agent, promises close when later evidence shows them done, and calendar moves settle the requests that caused them.
- 07
Each request gets what it needs
Team0 works out what the request is about, reads for it and hands back a short answer with sources. The agent doesn’t have to know which file to open.
- 08
You decide what each agent sees
Every agent has its own access. Stop one from saving while it keeps reading, or cut it off, without touching the others. Each read is logged.
Side by side.
| A markdown second brain | Team0 | |
|---|---|---|
| What fills it | What agents are told, and what they decide to write down | Your connected work, and what your agents save back |
| Who can read it | Agents that can open the folder or the repo | Every agent you connect, each with its own access, with no repo to share |
| When something changes | Overwritten, or both versions kept | Replaced, with the old version in history and the reason recorded |
| Heard versus decided | Every line weighs the same | Confirmed, heard, replaced and unknown are kept apart |
| Where it came from | Only if someone wrote it down | The exact source line on every statement |
| Two agents at once | One can overwrite the other’s edits | One writer, and every change is kept |
| Setting it up | Design the files, write instructions for each agent, give each agent access | Connect your sources, then connect each agent from Team0 |
| Upkeep | You prune, reconcile and check, and trust the agent with every update | Checks on every save and every night |
| Getting it to the agent | The agent loads an index and decides when to search | Team0 reads for the request and hands over what is relevant, with sources |
Where files still win.
We would rather you know this up front. Most people end up using both.
- Rules for how to work
- How to run the tests, the code style, what never to touch. A CLAUDE.md or AGENTS.md in the repo is the right home for that, and Team0 doesn’t replace it.
- No subscription
- Nothing to pay for and no service in the middle. The cost is your time instead: the setup and upkeep described above.
- Plain text you can open
- Any editor, any machine. In Team0 every statement shows its source, and you correct it once, in one place, for every agent.
Questions people with notes files ask.
- Do I have to give up my notes files?
- No. Keep CLAUDE.md and AGENTS.md for how your agents should work. Let Team0 carry what is true about your work, which is the part that keeps changing.
- My agents already write their own memory. Isn’t that the same?
- It is a good start, and it learns only from that agent’s own conversations. Team0 learns from your work as well, keeps track of what changed and who said it, and shares one understanding with every agent you use.
- What does Team0 keep from my agent conversations?
- It pulls out what matters and deletes the conversation itself after 7 days. What it learned stays, with the line each piece came from. What an agent says is kept as “the agent said”, never treated as fact on its own.
- Which agents work with it?
- Claude Code, Codex, OpenClaw and Hermes read and save on their own. ChatGPT, Claude, Claude Cowork, Town, monday.com, Base44 and other MCP agents read when you ask them to.
Keep reading.
- OverviewThe World Model
- ArchitectureFollow one fact from email to every agent
- Data modelStatements, two clocks, belief lanes
- Entity resolutionOne person, one record, carefully
- Truth maintenanceHow beliefs are retired, never deleted
- Trust and scopeWho said it, who may see it
- Understanding EngineFrom graph to a read an agent can use
- Memory vs understandingWhat a memory layer does, and what understanding adds
- ComparisonsSupermemory, Mem0, Letta, Zep, Claude and ChatGPT memory
- The 31 problemsEverything you have to solve, in one list
Your agents will change. Your understanding shouldn’t.
Invite-only private beta.