What it means to own the intelligence
3 November 2026
"Own your data" is an easy promise and most products that make it mean something narrow: you can download a zip file. Owning the intelligence is harder, and it is the part we think actually matters, so this is a plain account of what we mean by it and how Dutybound is built to deliver it.
The memory is a document, not a database you cannot see
Everything the service knows about you is written down in plain language, in a set of files you can open. Who you are. How you like things done. What you decided last month and why. Which vendors you distrust.
You can read it, edit any line, or strike a line out, and the model treats the edited version as the truth from then on. If you leave, this is what you take with you, and it is readable by a person or by any other model. We think this is the single most important design decision in the product, because it is what makes "the intelligence is yours" true rather than a slogan. A memory you cannot read is not yours.
The weekly report is the training signal, and it stays with you
Each week the service writes you a short report: what it did, what it learned about you, what it got wrong, and what it will do differently. You can push back on any of it.
That report is not a courtesy. It is the feedback the model is tuned on. In a consumer product, the equivalent signal is collected silently from millions of people and used to improve a model that belongs to the vendor. Here the signal is written to you, corrected by you, and used to tune something that belongs to you. The report also becomes a record of how your own judgment was applied over time, which turns out to be useful on its own.
Bring your model, run ours, or have one tuned to you
The base seat runs a frontier model through keys you hold, so the model vendor sees API traffic from your account, not from a shared pool, and your memory never leaves your environment to get there. You can switch models without losing anything, because the memory and the rules are not stored inside the model.
Members who want more can have a model fine-tuned on their own decisions, writing and standards. It is trained inside their environment and the weights are theirs. This is an add-on and it is not for everyone; the base configuration is good enough for most work, and we will say so when we think it is.
Approval rules are explicit, and the audit log is complete
You decide, per connection and per kind of action, what the service may do unasked, what it drafts for you to approve, and what it may never touch. It asks once, keeps the answer, and shows its reasoning when it acts. Every action taken on your behalf is written to an audit log you can read, with the instruction that caused it.
Storage is single-tenant, encrypted at rest with keys you can rotate, and members who want no shared surface at all can run on dedicated hardware or on their own.
People, under the same rules
Some work should not be done by software by itself: a phone call that needs judgement, or a signature. Members can add vetted human operators who work inside the same environment, under the same approval rules and the same confidentiality the model is held to. They do not get a copy of your data. They get a task, inside your system, with the permissions you set.
What is built and what is not
This is a beta that opens in January 2027, and we would rather be exact than impressive. The first version is being built in New York for the founder's own accounts and decisions, which is why the rules above are written the way they are: they are the rules he wanted for himself. The base seat, the readable memory, the approval rules, the audit log and the weekly report are what the beta opens with. Custom models and human operators come online during the beta, in that order, and founding members get first access to both. Private infrastructure is available from the start for members who ask for it.
We intend to publish the memory format before the beta opens, so that anyone can see what "a memory you can read" looks like before committing to anything. It is a folder of text files, and that is the point.
Dutybound opens a beta of 1,000 seats in January 2027 at $299 a month.