Every company we talk to has the same complaint in different words: “we set it up and the output is generic.” It is not the model. It is that the agent has never been told who the CEO is, what was decided in March, what shipped last week, or which customer is on the annual plan. That knowledge is the context layer, nobody ships it, and building it is most of the work of turning an AI employee into a digital teammate.
Ask ChatGPT to draft a reply to a customer and you get a polite, correct, useless paragraph. Ask Claude Code to implement “the thing three people asked for” and it asks you which thing. Ask a Hermes agent what is trending in support and it summarises today’s inbox with no memory of last month. None of these is a model failure. Each one is the agent doing exactly what a new hire would do on day one with no onboarding: sound professional and know nothing.
The difference between that and the drafts we actually send at Pocket Clear — the ones that know the feature shipped in 2.09, that this person wrote in from a different address than their account, that the last reply to them promised a fix — is not a better prompt. It is a layer of context the agent reads before it writes. Here is the pipeline; this post is about the layer underneath it.
A plain file per role. Not “you are a helpful assistant” — a specific person’s job: what they know, what they are responsible for, how they write, what they never say, who they escalate to. The CEO’s teammate and the HR teammate are different files with different sources, because a CEO’s digital teammate that can see payroll is a problem, and an HR teammate that speaks in the CEO’s investor voice is a different problem. What goes in each persona file, role by role.
The support inbox, the founder’s inbox for the CEO persona, the hiring inbox for the recruiter. Connected with read and draft scopes only — the agent reads mail and writes drafts into the same mailbox, and it cannot send. Every inbox is also a history: six months of past threads is the best onboarding document a teammate can have, and it is already written.
This is the piece teams either skip out of fear or do dangerously. The useful version is a view built for the agent: identifiers hashed or truncated, payment fields excluded, free-text columns stripped, no write path. The teammate can ask “is this account on the annual plan and when did it last sync?” and get a true answer without ever seeing a raw row. Document the masking rules; they are the first thing a reviewer will ask about. The general hardening is in the security checklist.
Decisions live in meetings. Record them — customer calls, standups, board calls, the Friday “what do we do about pricing” conversation — transcribe them, and index them so the agent can search “what did we decide about the Stripe migration?” and find the answer with a date and who said it. Then summarise each into a decision log: one line per decision, linked to the transcript. Without this, the teammate keeps proposing what you already rejected, because it was never told.
Want the context layer built for you?
Persona files, inboxes, sanitised data, transcripts, memory, plugins — set up per role and handed over with a runbook. From $1,000.
Most setups have session memory and nothing else, so every day is day one. Extended memory is boring and it is the whole game: a tally that grows (114 rows at Pocket Clear, 40 shipped), a contact log so nobody is emailed twice, a decision log, a “things I got wrong” file the human appends to when a draft is edited. All of it plain files in a workspace under version control, so it can be read, corrected, and rolled back. If the memory is in a vendor’s black box, you do not own the context layer.
Static context rots. The teammate needs a way to pull the current state of the things that change daily: the repo (what shipped, what is in review), the help centre (what article to link), the CRM (what stage this lead is at), the store reviews, the calendar. Skills, MCP servers, plugins — the name depends on the tool. The point is that “what is true right now” is fetched, not remembered.
A context layer that stores everything is a liability and, past a point, useless — the teammate cannot read a million tokens before every draft. So the rules: transcripts are kept verbatim but indexed, and only the decision log is loaded by default. Support emails are kept in the mailbox; only the one-line ask goes in the tally. Customer identifiers are never in memory files. Angry, legal and lost-data threads are flagged to a human and never summarised into a draft. The abstraction rules are the part of the layer that keeps it both safe and small.
The same tools produce different output. The CEO’s investor update arrives with this month’s real numbers and last month’s promises. The sales outreach references the objection the prospect raised on a call, because the call was indexed — and it is tuned on the threads that actually closed and the drafts you actually edited. The HR teammate answers a payroll question from the policy document and the pay calendar, not from the internet. Support drafts know what shipped. And once the layer exists, adding a job is cheap: the same teammate that reads the inbox can watch X, Reddit, Instagram and TikTok and hand you a weekly list of what to post — in your voice, because it knows your voice.
Because it is company-specific by definition. A model vendor cannot know your decision log; an agent framework cannot know your masking rules. Building the layer is a few days of unglamorous work — writing persona files with the people they represent, wiring read-only views, setting up transcription and indexing, deciding the abstraction rules — followed by a tuning window where drafts get edited and the memory gets corrected. It took us a weekend the first time. We now do it as a fixed-price AI Employee Setup: one persona from $1,000, the company layer from $10,000, priced here, on whichever pair of tools you already use — Hermes or Grok with Claude Code or ChatGPT.
The company-specific knowledge an agent reads before it writes: persona files per role, connected read-only inboxes, a sanitised read-only view of the user database, recorded and indexed meeting transcripts with a decision log, memory that extends (tally, contact log, corrections), plugins that fetch current state from the repo, help centre and CRM, and abstraction rules for what is summarised, kept, or never stored.
Because it knows nothing about your company. Claude Code, ChatGPT, Hermes and Grok all start from zero; without persona files, history, live data and memory, they produce correct, polite, useless output — like a new hire with no onboarding.
Yes, if the connection is a read-only view built for the agent with identifiers masked, payment and free-text fields excluded, and no write path — and the masking rules are documented. Never give an agent production credentials.
One persona with one inbox and one flow: two to three days. A company layer across six roles with database view, transcripts and repo: about two weeks, followed by a 30-day tuning window while drafts are edited and memory corrected.