Literess as an agent: the editor who remembers your decisions and does the work

This is a companion piece to our write-up on translation memory. That one was about what a translation tool should remember. This one is about what happens when the thing that remembers can also act — with the marketing turned down low.

There is a version of "AI in your translation tool" that everyone has seen by now: a chat panel pinned to the right edge of the screen. You can ask it to translate a paragraph, define a word, explain a grammar rule. It is genuinely useful, and it is also, structurally, a stranger. It does not know which project you are in. It does not remember that last week you decided this client's brand name stays untranslated. And when it suggests something good, you still have to be the one to go do it — copy it out, find the right place, paste it in, and hope you did not lose the thread.

We wanted the opposite of a stranger. We wanted an editor who has read the whole file, remembers the last three arguments you had about a term, and — when you ask — can actually make the change instead of just describing it.

That editor is Literess. This is the case for why she is an agent and not a chatbot, and what that word is doing here beyond being a fashionable label.

The difference between a chatbot and an agent

The word "agent" gets stretched to mean almost anything, so let us be precise about what we mean, because the distinction is the whole point.

A chatbot answers. You type a question, it produces text, and the loop ends there. The work of turning that text into a change in your document is yours.

An agent acts. It has tools, it can decide to use them, and it closes the loop between "here's what I'd do" and "it's done." The value is not that it is smarter — it is that it removes the manual carry between the suggestion and the result.

Both of those, on their own, are commodities in 2026. A chat panel over a translation is easy. A tool-using agent is, increasingly, also easy. What is not easy — and what we spent the real effort on — is making the agent's actions worth trusting. An agent that acts confidently on the wrong context is worse than no agent at all, because now it is generating mistakes at machine speed and you are the one cleaning up after it.

So the two things we care about are, in order:

1. What does she know? — the memory she is grounded in.
2. What can she do with it? — the actions, and the guardrails around them.

What Literess knows: decision context, not just words

If you have read the translation-memory piece, the foundation will be familiar. If you have not, here is the short version.

Most translation memory remembers the sentence — the source text, the approved translation, and some metadata. That is table stakes. What is rare, and what Transept is built around, is remembering the work that made the sentence trustworthy: the alternatives that were rejected, the comments in the margins, the review history, the reasoning behind why one version won over another. We call it decision context, and the important move is that we feed it back into the model so the next translation can reuse the reasoning, not just copy the output.

Literess drinks from exactly this well. She is not a general assistant who happens to be sitting inside a translation tool. She is grounded in:

• your projects and the documents in them,
• your glossaries and styleguides — the terms and voice you have already committed to,
• past decisions — the approved translations and the discussion that produced them, including the alternatives you turned down,
• and the live document you are working in right now.

This is why she does not feel like a stranger. When she suggests a phrasing, it is a phrasing shaped by what you have already approved elsewhere. When she flags a term, she can tell you not just "this looks off" but "you translated this differently in the last two documents, and here is the comment where you explained why." She is reading from the same memory the whole product runs on — the reasoning, not just the words.

An editor who remembers your decisions is worth a great deal on its own. But memory that can only speak is still only half of what a good editor does. The other half is doing the work.

What Literess does: real actions, always confirmed

Here is the part that makes "agent" the honest word.

Literess does not stop at description. When you ask her to do something, and you confirm it, she does it — inside the product, with the same tools a human on your team would use. In practice that means she can:

• Create and run workflows — the multi-step review passes (translate, then proofread, then check against the styleguide, and so on) that Transept automates. She can assemble one for the task at hand and set it running.
• Draft documents — start a new document for you rather than handing you text to paste somewhere yourself.
• Move work across the CRM board — nudge a project from one status to the next as it progresses, so the board reflects reality without you switching contexts.
• Set defaults — make a particular glossary or styleguide the default for a document, so the right memory is in play from the first line.
• Invite teammates — bring a collaborator into the work when the job needs another pair of hands.
• Drive the editor itself — open the panel you need, navigate you to the right screen, put you where the work is instead of describing where to click.

Read that list again with the emphasis on the verbs. These are not answers about how to do things. They are the things, done.

And every one of them is gated behind your confirmation. This is not a footnote — it is a design principle, and it is the difference between an assistant and a loose cannon. When Literess proposes an action, it surfaces as a small confirmation chip: here is what I'm about to do, do you want me to? Nothing happens to your projects, your documents, or your board until you say yes. She acts on your behalf, never behind your back.

We were deliberate about this. It would have been easy to make her more autonomous — to let her fire off workflows and reorganize your board because she was fairly sure that is what you wanted. We chose not to. In translation, the cost of a wrong-but-fast decision is high and quiet: a term drifts, a voice slips, a client's copyrighted phrasing ends up somewhere it should not. The confirmation step is cheap. The mistake it prevents is not.

Why "remembers and acts" changes the shape of the work

Put the two halves together and something shifts that neither half delivers alone.

A tool that only remembers still leaves you as the sole operator. You are the one who has to notice the relevant past decision, go find it, and apply it. The memory is a filing cabinet — excellent, well-organized, and entirely passive.

A tool that only acts — a tool-using agent with no real grounding — is fast and confidently wrong. It will happily run a workflow with the wrong glossary because it does not actually know which one you meant.

An editor that does both turns translation from a static artifact into a living workspace. The classic mental model of machine translation is a slab: the machine hands you a wall of text, and your job is to walk along it fixing what is broken. It is finished the moment it arrives, and it forgets everything the moment you close it.

What we are building is the opposite of a slab. Every decision you make feeds the memory. The memory grounds Literess. Literess, with your confirmation, does the mechanical work of applying those decisions — running the passes, setting the defaults, keeping the board honest — so your attention stays on the part only you can do: deciding what sounds right. The next document does not start from zero; it starts from everything the last one taught the system. The workspace compounds.

That is the thesis in one line: an editor that remembers and acts changes translation from a slab you fix into a workspace that gets better the more you use it.

What she is not

We should be as clear about the limits as about the claims, because overselling this is the fastest way to lose your trust.

Literess is a literary editor, not a replacement for human judgment. She runs on frontier models — currently Google's Gemini — and those models are extraordinary at the mechanical breadth of the job: consistency across a 40,000-word manuscript, remembering a term from chapter two, catching the drift you would have missed on the fortieth page. They are not the person who decides that a joke should stay a little awkward because the awkwardness is the point. That call is yours, and it should be.

She does not act autonomously. Every action she takes, you approved. The quality mode you choose — Fast, Standard, or Pro — is your decision. Which of her suggestions become the final text is your decision. She is very good at making sure nothing is forgotten; she is not in the business of deciding what is good. That distinction is not a limitation we are apologizing for. It is the whole design.

An editor who remembers everything and does the tedious work, so the human can spend their attention on the choices that actually need a human — that is what we were trying to build. Not a chatbot on the side. An agent in the room.