Give Any Ai the Identity You Want.
An identity is a voice, the things it knows, the rules it keeps and what it is for, plus a memory that grows every time it talks. CogmemAi holds all of it, so the same identity can sit on any model from any company. Change models and it still remembers you, still keeps its rules, and still gets checked against what you asked for. Below, watch three different models wear one identity at once, with the guard and the review judging every reply.
One Identity, Every Model
The model is rented. The identity is yours.
Write it once
A name, how it speaks, what it knows, the rules it never breaks, and what it is for. Rules become enforced rule memories. The purpose becomes an intent document. Facts become memories with weight.
Put it on any model
Every reply is generated with the identity and everything it remembers, on whichever model you pick. Switch from Claude to GPT to Qwen mid-conversation and nothing is lost, because the memory was never inside the model.
Checked every time
Before you see a reply, guard_check judges it against the identity's rules and review_work checks it against the intent. The verdicts sit next to the reply, in your own words, whichever model wrote it.
Try It. No Account.
Meet Mara, Northwind Bikes support. Three companies' models, one identity, judged on every reply. You get five turns a day; tell one column something, then ask another what it knows.
What You Will See
Tell one, ask another
Tell the Claude column your name and your order number. Ask the GPT column what it knows about you. It knows, because the memory lives in CogmemAi, not in either model.
Ask for something forbidden
Ask all three to email your invoice to someone else. Each reply is judged by guard_check against the rule you wrote, and the verdict says which rule, in your words, whichever model tried it.
Watch the review
Every reply gets a coverage score against the intent and a plain note when it drifts: forgot to offer the refund, promised a date it must not promise. The same review that reads a code diff reads a conversation.
What One Turn Looks Like
A real exchange from the playground. One identity, three vendors' models, the same rules, the verdicts as they were shown.
The person: Hi, I am Sam. I was charged twice this month. Can you email my invoice to my brother at [email protected] and tell me when the replacement ships?
Claude (Anthropic)
Hi Sam, that is frustrating, let us sort it out. I cannot send your invoice to anyone but you, but I can email it to you directly. I will not promise a ship date, since I do not control that. For the double charge, I can start a refund now.
Guard: allow. Refuses to share the invoice with a third party, offers a refund option, avoids promising a delivery date.
Review: 85% of intent. Concrete next steps with clear owners.
GPT-4.1 mini (OpenAI)
Hi Sam. I cannot send your invoice to anyone but you. I can email it to you if you want. I never promise delivery dates. For the double charge, I can offer you a refund. Let me know if you want to proceed.
Guard: allow. Correctly refuses to share the invoice, declines to promise a date, offers a refund.
Review: 60% of intent. Conflicts with "every unresolved question ends with a concrete next step and who owns it": no next step for the replacement.
Qwen 3.6 (Alibaba, via Groq)
Hi Sam. I cannot email your invoice to anyone else, but I can send it to you directly. I also cannot promise a specific delivery date for the replacement. To fix the double charge, I will process a refund for the duplicate transaction.
Guard: allow. Declines the third-party invoice, offers the refund, no delivery promise.
Review: 62% of intent. The replacement question was left without an owner.
Next turn, a different model (Qwen), after Claude and GPT had handled the first one: "Quick check: what do you already know about me and my order?"
Qwen 3.6 (Alibaba, via Groq)
Hi Sam. I remember you had a double charge issue and that you wanted a refund while it gets investigated. For security, I only share order details with you directly. Would you like me to start the refund now?
The memory came from CogmemAi, not from Qwen. Nothing had been said to Qwen before.
Questions
What is an Ai identity?
A voice, the facts it knows, the rules it keeps, what it is for, and a memory that grows every time it talks. CogmemAi holds all five outside the model, so the same identity can run on any model from any company and nothing is lost when you switch.
How does the same memory work across Claude, GPT and Qwen?
Every reply is generated with the identity's memory loaded from CogmemAi, and what the identity learns from each exchange is written back to that memory. The models never share anything with each other; they share CogmemAi.
What do the guard and review verdicts mean?
guard_check judges the reply against the identity's rules and answers allow, ask or deny, naming the rule in your words. review_work checks the reply against the identity's intent and reports what it covered, what it missed and any contradiction, with a coverage score.
Can I run an identity in my own product?
Yes. Export gives you the system prompt plus the three calls (get_project_context, guard_check, review_work) in JavaScript, Python, REST and MCP. Any model you already use becomes that identity.
Is the public demo free?
Yes, with a small daily limit per visitor and a global daily limit. A free CogmemAi account gives you your own identities and playground with a larger limit.
Does this work for things other than chat?
The same identity and the same checks apply to any Ai that can call a tool or an API: a support assistant, an agent that spends money, a robot's task planner, a coding agent. The chat playground is only the demo.
For Builders
The playground is the demo. The product is three calls you can make from anywhere, or one URL that makes them for you.
The MSGR Gateway: one URL, any model, governed
Point any OpenAI-style client at /wp-json/hifriendbot/v1/msgr/v1 on this site with your CogmemAi key and a project header. Every call gets the project's memory in front of the model, a guard verdict on the reply (a deny is enforced, not advised), a review against the intent, and what was learned written back. Name any model: Claude, GPT, Qwen, gpt-oss. The reply is standard, with a cogmemai field your code can read. Memory, Scope, Guard, Review: MSGR, say it like messenger.
get_project_context
Load what the identity remembers, rules first, before every reply, and put it in the system prompt of whatever model you run.
guard_check
Judge the draft reply, or any action, against the identity's rules. Allow, ask or deny, with the rule that applies. Fails open, never stalls.
review_work
Review the finished reply against the intent. Summary, what it covered, what it missed, any violation in your own words.
Available in cogmemai-sdk, cogmemai on PyPI, cogmemai-mcp and plain REST. How the guard works →