How the memory works,from start to finish.
Whichever AI you use, it reads the same memory, takes corrections in a single sentence, and picks up where you left off.
Capture it, keep it,
read it, refine it.
Pick what's worth keeping, put it in the canonical, read just what's needed, fix it with a sentence, and round again.
OTHER AI MEMORYThe model decides what to remember and when to change it; people review afterwards.
moAIPeople decide what matters; the AI does the editing.
See detailsClose
- 01CAPTURE
Capture
From chats and meetings, pick only what should last: decisions, status changes, procedures. Whole conversations are not stored.
- 02CANONICAL
Canonical
What you pick goes into the canonical documents once a person confirms it. The canonical is the memory.
- 03READ
Read
Connected AIs take only the lines they need from the canonical, together with the version they read.
- 04WRITE
Refine
Ask in plain words and the AI makes the change. It shows the change first, never writes over a document that changed in the meantime, and logs the result.
ElsewhereMost AI memory runs these four steps on its own. The model decides what to remember and when to change it; people review or delete afterwards. Memory in ChatGPT, Claude and Gemini, and memory layers that automatically extract facts from chats, work this way.
In moAIIn moAI, the more important a memory is, the more it's decided by people, while the AI does the editing, and every AI works from that same standard.
Your AI's memory,
as documents you can read.
Memory lives not as fragments in a database, but in documents people can open and read: the canonical.
OTHER AI MEMORYMany systems store short fact sentences pulled from chats in a database.
moAImoAI keeps each decision and its reason together, in a document people read.
See detailsClose
Status
Preparing the October workshop. Venue confirmed; second review of the materials under way.
Decisions
10/02 · The workshop runs 90 minutes, hands-on.
Why: hands-on practice was the top request in the last survey
Next
Final materials · Note to participants
Change log
10/02 format confirmed · 09/28 venue changed
Older entries moved to the archive
- APeople and AI read the same page
- BEach decision keeps its reason and date
- COld entries are archived word for word
Open it, and you know what your AI remembers.
It's not a summary screen; it's the very text your AI reads. If something's wrong, correct it to your AI in a sentence, or fix it right there yourself.
Decisions keep their reasons.
Not just what was decided, but why and when it changed, all in one place, so the thinking can be retraced later.
Only what matters, kept light.
Status, decisions, next steps and a change log make up the basic frame. Older entries move, word for word, to an archive so the canonical stays light.
ElsewhereMany memory systems pull short fact sentences out of chats and store them in a database (for example, mem0 and Zep). They build up quickly, but the reasoning behind a fact, such as why it was decided, tends to stay outside the short sentence.
In moAIKeeping memory in human-readable files is already common in developer tools (for example, CLAUDE.md in Claude Code). moAI extends that approach to every AI and the whole team, and adds permissions and version checks.
Tools may change;
the standard stays in one place.
Keep using ChatGPT, Claude, Codex or Hermes; each reads the same canonical through MCP, a standard connection.
OTHER AI MEMORYEach service's memory builds up inside that service's account.
moAImoAI uses a standard for the connection and keeps the memory on your storage.
See detailsClose
- ChatGPTOwn memory stays
- ClaudeOwn memory stays
- CodexOwn memory stays
- HermesOwn memory stays
Start guide
- Shared rules
- Document map
- Canonical by topic
- What may be edited
Same identity · same permissions
Every AI starts from the same guide.
At the start of a chat, each AI reads a start guide: which documents to read first and what it may change. No more repeating yourself for every tool.
Permissions follow people and projects, not tools.
Whichever AI connects, the same identity and permissions apply. A document you can't see stays hidden in every tool.
Each AI keeps its own memory.
Convenience memory stays inside each tool; lasting standards live in the canonical. When they disagree, the canonical wins.
ElsewhereMemory in ChatGPT, Claude and Gemini builds up inside each service's account; moving it elsewhere usually means exporting and pasting. Some memory services do connect many tools over MCP, but the memory usually lives in that service's cloud account (for example, Zep and Supermemory).
In moAImoAI uses a standard for the connection and keeps the memory on your storage.
Beyond ChatGPT on its own:
read, fix, carry on.
On its own, ChatGPT can't edit your documents. Connect the moAI app, and even a new chat on the web or desktop reads the same canonical and takes corrections in a sentence.
CHATGPT MEMORYChatGPT's memory builds up and is used inside your ChatGPT account.
moAImoAI's canonical is read and edited by ChatGPT and your other AIs alike.
Easy to mix up. Here's how they differ.
- CHATGPT
An AI you talk to
It reads files you add to a chat or project. But the apps OpenAI builds in can only find and read material, so it can't edit your original documents.
- CODEX
A coding agent for developers
A separate tool included in paid ChatGPT plans. It reads, edits and runs code in a working folder you choose, mostly for development work.
- CHATGPT + moAI
Read and edit the canonical in the chat you already use
No developer tools needed. Fix things with a sentence, and gain what ChatGPT didn't have: version checks, permissions and work that carries over.
Per OpenAI's documentation (checked 2026-10-08). Codex can connect to the same canonical over MCP, too.
See detailsClose
Web and desktop, one connection.
Publish the moAI app to your workspace once, and members find and read the canonical through the same app on the ChatGPT web and desktop app.
Fix it right in the chat.
Say “This decision changed,” and you see the change first; it reaches the canonical only after the checks.
A new chat, the same context.
Open a fresh chat and the canonical and task records are still there, so there's no explaining from scratch.
According to OpenAI, MCP apps with write actions are in beta for ChatGPT Business, Enterprise and Edu workspaces and are not yet available in the mobile app. Scope varies by plan and permissions.
STATUSPilot-verified · in the ChatGPT desktop app: create, read, partial edit and read-back of a document
New AIs keep coming;
they join the same memory.
Any AI that plugs in tools over MCP, such as Aside, Buzz, OpenClaw or NVIDIA NemoClaw, can join the same canonical, so there's no fixed limit to what you connect.
TOOL BY TOOLEach tool keeps its own memory, so more tools means more scattered memory.
moAIWith moAI, tools can multiply while memory stays in one place.
See detailsClose
- Aside
AI browser · finds, reads and edits the canonical with moAI tools
Pilot-verified - Buzz
Block's agent platform · installed so agents look to the moAI canonical first
Installed · reads being verified - OpenClaw
Open-source agent · supports MCP servers, per its documentation
Official route · moAI link not yet verified - NVIDIA NemoClaw
NVIDIA's open-source stack that runs agents such as OpenClaw and Hermes in OpenShell sandboxes · connects through the agents inside
Official route · moAI link not yet verified - The next tool
If it supports MCP, it connects the same way. Actual support is checked before connecting.
Checked before connecting
Based on each official documentation as of October 8, 2026. Product names and trademarks belong to their owners and imply no affiliation or endorsement.
Not reading more,
but reading exactly.
It finds the relevant documents, reads only the lines it needs, and notes which version and which lines it read.
OTHER AI MEMORYWhen similar snippets are gathered, old and new versions of a fact can surface together.
moAImoAI reads the current version of the edited canonical.
See detailsClose
- 01 / FINDFind the documents
- 02 / READ LINESRead only the lines needed
- 03 / CITEKeep the source and version
How long is this workshop again?
90 minutes, hands-on.
workshop.md · Decisions · lines 12–14 · REV 7F3A
No unread ranges
What wasn't read is never hidden.
If something was off-limits or cut off, the answer says so. Gaps aren't filled with guesses.
The relationship map only points the way.
The Graphify map narrows down which documents to read. Facts are always checked in the original.
The original beats the summary.
If a quick summary or an AI's own memory disagrees with the original, moAI goes back to the original.
ElsewhereSimilarity search quickly gathers snippets that resemble the question. But when old and new versions of the same fact are both stored, both come up, and search ranking ends up deciding which is current. It happens often with extraction-based memory that only ever adds entries.
In moAImoAI reads the current version of the edited canonical. The document itself tells you which decision is current.
Less to read,
faster and sharper answers.
Code does the finding and sorting first, and the AI gets only the context it truly needs.
WHOLE DOCUMENTSFeeding in whole documents means more to read, slower answers and a blurrier focus.
moAImoAI passes along only the lines the question needs, with their source.
See detailsClose
Code finds and sorts first.
Repetitive steps such as searching, filtering and bundling run as code in one go, instead of the AI asking step by step. Only the result goes back to the AI.
Just the lines needed, with sources.
Context is bundled within set ranges and size limits, and anything cut off or unread is flagged.
So all three improve together.
Less to read means fewer tokens, shorter waits, and clearer grounding for more accurate answers.
STATUSIn use in LaB96's own operations · effects are measured per setup and not guaranteed
Fix it with a sentence;
it's checked three times.
Say “It's 90 minutes now,” and the AI finds the line and fixes it, then confirms with a preview, a version check and a read-back.
OTHER AI MEMORYWhen an agent changes its memory on its own, you often learn only later that a standard changed.
moAIIn moAI the AI does the editing too, but it shows the change first and logs it.
See detailsClose
- 01 / PREVIEW
Preview
−60 minutes, lecture
+90 minutes, hands-on
- 02 / CHECK
Version check
Version read = current version
7F3A = 7F3A ✓
- 03 / APPLY
Apply
Once, by request number
REQ #A1C9
- 04 / READ BACK
Read back
Matches exactly
✓ EXACT
- 05 / LOG
Log
Change log +1
REV 7F3A → 2B90
If someone changed it in the meantime moAI stops instead of overwriting, rereads the latest and proposes again.
No silent overwrites.
Even when two AIs or two people edit the same document, a change made without seeing the newer version is refused.
Or edit it yourself.
You can also open the document and edit it directly. The next AI to read it checks and reads the new version.
Only what people confirm becomes canonical.
Memory candidates from chats don't become canonical on their own. Only confirmed decisions go in.
ElsewhereIn many systems the AI edits memory on its own: an agent rewrites its own memory blocks (for example, Letta), or memory items are added automatically during chats (for example, ChatGPT and Claude). It's convenient, but you often find out only later that an important standard changed.
In moAIIn moAI the AI does the editing too. But it shows the change first, checks the version before applying it, and logs it, so the moment a standard changes is never missed.
Open a new chat
without starting over.
Goals, done criteria and progress live in a task record, so another AI or device picks up right where you left off.
OTHER AI MEMORYApp memory and projects carry on inside the same app.
moAImoAI's task records live outside the tools, so it's the same task whichever AI opens it.
See detailsClose
Finish the workshop materials
Done when
- 3 hands-on exercises
- A 90-minute schedule
- A note to participants
- MON · MAC · CODEXTask started
- WED · WEB · CHATGPTProgress check 2/3
- THU · MAC · CLAUDEPicks up and finishes
- Ran
- Checked
- Delivered
- Confirmed by a person
Done and confirmed are recorded separately.
What finished running, what was checked, what was delivered and what a person approved are each kept apart.
When unsure, check first.
If an outcome is uncertain, the real state is checked before anything runs again. Nothing is done twice.
One task, one owner.
One place runs each task, so several tools don't each repeat the same goal.
ElsewhereApp memory and project features help the next chat inside the same app. But they rarely carry a task's done criteria and progress over to another app or another AI.
In moAImoAI's task records live outside the tools, so it's the same task whichever AI opens it.
Who sees what
is up to you.
Memory stays on your storage by default, personal, team and sensitive areas are kept apart, and anything irreversible asks first.
OTHER AI MEMORYA lot of AI memory is stored in the provider's cloud account.
moAImoAI's canonical is simply documents you own.
See detailsClose
Memory · documents · search index · task records
Accounts, registered devices and allowed setups only
- Sending out
- Deleting
- Payments
- Schedule changes
See it, edit it, export it, delete it.
The canonical is ordinary documents that open without any particular service. You can manage it yourself, any time.
Irreversible actions wait for approval.
Sending things out, deleting, paying or changing schedules never happens without explicit approval.
The AI subscriptions you already have.
AI calls go through the subscriptions you already use and never switch to another billing route on their own.
ElsewhereA lot of AI memory is stored in the provider's cloud account and managed through settings screens or chat (for example, ChatGPT, Claude, Gemini and managed memory services). Some open-source options can be self-hosted, but often the memory still lives as database records.
In moAImoAI's canonical is simply documents you own.
One team's week,
day by day.
Here's how one decision becomes canonical, and how several AIs and people then work from it.
- MONA meeting decision becomes canonical
- TUEClaude reads it and drafts
- WEDChatGPT gives the same answer
- THUA new teammate, one document
- FRIChanged in a sentence, logged
See detailsClose
- MON
In a meeting, the team settles on a 90-minute, hands-on workshop. It's suggested as a memory candidate, the lead confirms it, and it goes into the canonical.
- TUE
While drafting materials with Claude, Claude reads the same decision from the canonical and fits the plan to 90 minutes.
- WED
Out of the office, a teammate asks ChatGPT on the web, “How long was it?” The answer: “90 minutes, hands-on. Decided 10/02.”
- THU
A new teammate joins and gets the status, decisions and next steps from a single canonical document.
- FRI
The venue forces a schedule change. One sentence, “Move it to Friday afternoon,” and the AI previews the change and applies it once confirmed. The reason stays in the change log.
A synthetic example for illustration. Actual setups and connected tools depend on each customer's environment.
Next to other AI memory,
at a glance.
Each approach has its strengths. moAI works alongside these tools and takes care of the standards that need to last.
| moAI | App memoryChatGPT · Claude · Gemini | Extraction-based memorymem0 · Zep | Agent memoryLetta | |
|---|---|---|---|---|
| What memory is | Canonical documents people read | Memory items and summaries | Fact sentences · graph records | Per-agent memory blocks · files |
| Who decides what's kept | Decisions and status people confirm | The AI, during chats | An LLM extraction step | The agent itself |
| What counts as current | The edited text and its change log | Updated by the service | Search ranking or validity dates | Edited by the agent |
| Across AIs and tools | Every AI connected over MCP, including ChatGPT on the web and desktop | Within that service | Via API or MCP | Centered on that agent |
| Where it's kept | Your devices and storage (by default) | The service account | Managed cloud or self-hosted | Cloud or local |
| Before a change lands | Preview · version check · read-back | Edit or delete afterwards | Edit or delete via API, with event history | A record per edit (e.g. git commits) |
See detailsClose
Based on each service's official documentation as of October 8, 2026. Details vary by plan, region and version. Product names and trademarks belong to their owners, and this comparison implies no affiliation or endorsement.
Start with what you need,
in the smallest setup.
Start with the memory foundation, and add features only when a real bottleneck appears.
See detailsClose
- BASE
Memory base
Canonical · start guide · ownership and permission rules
Core - MAP
Relationship map
Explore how documents and decisions connect, with Graphify
Optional - DIGEST
Status digest
A quick-read summary, checked against the canonical when stale
Optional - CAPTURE
Capture suggestions
Candidates drawn from chats and meetings, approved by people
Optional - PARTNER
Always-on partner
An always-on assistant for long-term curation and briefings
Optional - ACTION
Action automation
Only once approval, records and rollback are in place
Optional
Questions come before tools.
First we ask what you keep re-explaining, which decisions must still be remembered a month from now, who should see what, and what must never be seen.
The higher the stakes, the slower we go.
With sensitive material or costly mistakes, automatic capture and automatic actions stay off by default.
Results are measured.
How often you re-explain, how long it takes to find a decision, how long until it's recorded: real work metrics decide the next step.
Not just delivered,
but fitted to your expertise.
The standard setup works on its own; if you like, we'll also design and run AI solutions fitted to your work and expertise, as a consulting option.
See detailsClose
We map your field together.
Interviews uncover where your expertise goes into decisions and what you keep explaining.
Your standards go into the canonical.
Terms, quality bars and review practices go into the canonical and start guide, so every AI works to the same standard.
Solutions built for your work.
We design the flow of repeat work, review steps and the automation you need, and connect them only where approvals and records are in place.
We keep refining as you go.
Monthly operations keep the canonical light, and new AI tools and features are added as their real effect shows.
Consulting is optional. You can choose the standard setup alone, and what's included and what it costs are spelled out before any contract.
Frequently asked
questions.
Do I need to turn off my AI's own memory?
No. Each tool keeps its own memory; only the standards that need to last go into the canonical. When the two disagree, the canonical wins.
What if the AI remembers something wrong?
Just correct it in a sentence. Say something like “The workshop is 90 minutes now,” and the AI finds the relevant part of the canonical, fixes it down to the exact wording, and shows you the change before applying it. You can of course edit the document yourself, too. Either way, every connected AI reads the corrected version from then on.
Does it work in ChatGPT?
Yes. Through the moAI app you can read and edit the canonical in ChatGPT on the web and in the desktop app. Per OpenAI, using both reading and editing requires a ChatGPT Business, Enterprise or Edu workspace, and the mobile app isn't supported yet.
How is ChatGPT different from Codex?
ChatGPT is an AI you talk to; Codex is a coding agent for developers, included in paid ChatGPT plans, that reads, edits and runs code in a folder you choose. The apps OpenAI builds into ChatGPT can find and read material but can't edit original documents. With moAI connected, you can read and edit the canonical and carry work on right in the ChatGPT chat, no developer tools needed. Codex can also connect to the same canonical over MCP.
Are all my conversations saved?
No. Only things worth keeping, such as decisions, status changes and procedures, become candidates, and only confirmed ones go into the canonical. Automatic capture is optional and starts with suggestions only.
Do I have to use Obsidian?
No. Any documents people can read and edit will do. Obsidian is one convenient default.
Can I use it if I'm not a developer?
Yes. We set up and connect everything with you, and show you how to ask and correct in plain words and how to check the results.
How much does it cost?
There's an initial fee for diagnosis, design and setup, a monthly operations fee, and the cost of your AI subscriptions and any extra infrastructure. Consulting fitted to your expertise is optional, and what's included and what's extra is spelled out before any contract.
Start with
one piece of work.
One thing you keep explaining,
or one project you want to keep going, is enough.
Your mail app will open. This page does not collect any inquiry data.