Per-Client Glossary Memory for Freelance Translators with AI

The 45 Minutes You Lose Every Monday Morning

Every freelance translator starts Monday the same way. Open the CAT tool. Open last week's project folder. Pull up the client's style guide PDF. Find the glossary spreadsheet. Remember that Client A prefers "utilisateur" but Client B insists on "utilisatrice" as the neutral form. Remember that Client C's brand voice is formal French with no contractions, while Client D runs informal voseo. Remember that in the gaming project, "weapon" stays in English while "armor" gets translated. Remember.

By the time you are actually ready to translate, you have burned 45 minutes just reloading mental state. Multiply by four or five clients in a normal freelance week, and that is three hours of pure context-loading time, before you charge a cent.

The real solution is not yet another glossary spreadsheet, another terminology database, or another memoQ termbase export. The real solution is an AI agent that remembers per-client context for you and is accessible from wherever you are working. A Hermes agent on Telegram with per-client glossary skills is the cheapest way to get that today, and it pays for itself the first Monday you skip the 45-minute warmup.

What Client Memory Should Actually Hold

Most translators already know the theory. In practice, per-client memory needs to hold a handful of specific things:

Terminology glossary with source term, preferred target term, and any do-not-translate rules.
Style preferences. Formal or informal register, contractions allowed or not, sentence length tolerance, heading capitalization.
Tone. Brand voice described in plain language with three or four example sentences.
Past decisions. "We discussed and agreed to translate 'sustainability' as 'sostenibilità' not 'durabilità' after a call in September."
Client quirks. "Always use the Oxford comma for this client. They will return files without it."
Reference files already delivered, so the agent can pull prior phrasing.
Do-not-translate list. Brand names, product names, specific acronyms.

That is seven categories per client. Maintaining them manually across 5 to 20 clients is the reason most translators have a half-updated termbase and a Google Doc with notes that never get opened.

How an AI Agent Replaces the Ritual

A Telegram AI agent with per-client skills changes the ritual from "load context manually" to "ask the agent." Examples of actual queries a working translator sends the agent during the day:

"For Client A, how do we translate 'user experience'?"
"What tone am I writing for Client C, show me examples."
"Is 'Platform' capitalized in Client B's deliverables?"
"I just got this string from Client D, compare it to their style guide and flag anything off."
"Add 'geofencing' to Client A's glossary, target is 'geolocalizzazione' per today's client email."
"Pull the last three times I translated 'onboarding' for Client B and show me."
"Client E sent a new 2-page style update, parse it and add the changes to their memory."

The agent answers in seconds from memory. You never open the PDF. You never search the spreadsheet. You ask, you get, you translate.

How the Glossary Gets Built

One evening per client. That is the whole setup.

Step 1, paste existing assets. Your current glossary CSV, the client's style guide PDF, past delivered files if you have them. The agent reads and normalizes all of it into its own per-client skill.

Step 2, write a short free-text brief. Two paragraphs about the client's voice and any quirks that never make it into official docs. "They never accept passive voice in marketing copy, they are fine with it in documentation. They have a running joke about never using the word 'leverage.' Do not use it."

Step 3, tell the agent your workflow. "When I ask about Client A in Telegram, always load their glossary skill. When I paste a source segment with '#A' at the start, treat it as belonging to Client A."

That is it. You repeat for each active client. Over a few weeks the glossary grows organically from your real daily queries.

Real Daily Use Inside a CAT Tool Workflow

You are still working in Trados, memoQ, Wordfast, or Phrase. The agent does not replace the CAT tool. It sits next to it in a different app on your second screen or on your phone.

9:00am. You open Monday's project for Client A. Before you start, you type in Telegram: "#A what is our latest glossary status?" Agent replies with the top 20 terms and any changes since last Monday.

9:17am. Source segment reads "implement the new geofencing feature." You are not sure if geofencing has been locked in. You type "#A geofencing." Agent: "Locked last Thursday as 'geolocalizzazione', source email attached. Client explicitly rejected 'geofencing' as an English loanword."

9:45am. You hit a segment you are unsure about tone for. You paste it into Telegram: "#A is this in their voice?" Agent reads the segment, compares to style guide, returns: "Slightly too formal. Consider using 'ti aiutiamo a' instead of 'vi permettiamo di'. Matches their tone from the March newsletter."

11:30am. Client emails a revised term. You forward the email to the agent: "#A add this to glossary." Agent reads, updates, confirms. No extra spreadsheet edit.

Saturday afternoon. You are halfway through a new project for Client A. Mid-session, you realize a term rings a bell. You type "#A have we seen 'predictive scheduling' before?" Agent: "Yes, project from March 14. You translated it as 'pianificazione predittiva'. Source email and final delivered segment attached."

Why This Beats a Shared Termbase

Shared termbases and tools like memoQ's QTerm or Trados' MultiTerm exist for exactly this problem. They work. They are also rigid. Three things they do not do that an agent does.

Ask in plain language. You do not query memoQ in English. You click through a dialogue. An agent takes "what's the tone for Client C" and gives you a paragraph of answer. Faster on the margin, dozens of times a day.

Explain the why. A termbase tells you the term. An agent tells you the term plus the reason: "Client B requires 'platform' with lowercase because their brand lead wrote it in the 2024 style update." The why matters when you have to defend a choice.

Remember decisions that never made it to the termbase. Most translators do not add every decision to their termbase because it is friction. They do send Telegram messages to themselves. An agent with persistent memory in Telegram captures those decisions for free because they happen in the channel the agent already owns.

What the Agent Does Not Replace

Three things stay with you.

The CAT tool itself. Segmentation, TM matching, QA checks, delivery format, these are what CAT tools are built for. Use them.

Client relationships. The five-minute call to lock a tricky term, the email explaining why you chose a phrasing. Those are human and always will be.

Final judgment. The agent's glossary recommendations are a starting point. If a segment calls for breaking a rule, you break it. You are the translator.

The Economics for a Working Freelancer

Numbers that matter. A full-time freelance translator working 6 hours a day on actual translation plus 2 hours of context-loading, QA, admin, and client communication is leaving 2 paid hours per day on the table. Reducing context-loading by even 30 minutes a day through per-client memory frees up time worth roughly 150 euros or dollars at an average 0.10-0.12 per word rate and normal speed.

A Hermes agent on Hermify costs 12 dollars per month plus your own LLM API usage, typically a few dollars per month for this kind of usage. One reclaimed hour a week pays for the whole year. Ten reclaimed hours a month pays for a nice vacation.

The Part About Confidentiality

Translation work often comes with NDAs. Fair concern. Hermes Agent runs on your own LLM API key, which means API calls go directly from your agent to your provider, not through a third-party SaaS. Your client data stays between you, your agent, and your LLM provider (OpenAI, Anthropic, etc.) per their own data policies. If you work with clients who specifically require local models, you can run Hermes with an open model and self-hosted endpoint, which is a harder setup but possible.

Either way, you end up with a cleaner privacy story than "I pasted the client doc into ChatGPT to ask a question," which is what most freelancers quietly do.

Setting Up This Weekend

Spin up a Hermes agent on Hermify. Sixty seconds.
Pick your top 3 clients. Load their glossary, style guide, voice examples, and any quirks.
Set up the hashtag shorthand. "#A", "#B", "#C" for fast context switching.
Use the agent for one full working week. At the end of the week, review what you asked the agent about most. Those topics become new skills or glossary entries.
Add the remaining clients over the next two weeks. Do it lazily: a client gets added when their next project comes in, not in an upfront batch.

Within a month you will not remember how you worked without it.