Two agency consultants reviewing documents from several clients at a meeting table

Use case: an agency with 12 clients that remembers what it promised to whom

7 min read

A simulated scenario: an agency with 12 clients, promises scattered across emails and calls, and what client separation MemoryFirst offers and what it does not.

by Equipo MemoryFirst

Table of contents
  1. When the promise lives in an email from six weeks ago
  2. What would change with and without memory, with illustrative numbers
  3. How what was promised gets recorded
  4. Separation between clients: what exists and what does not
  5. What to choose depending on what you have been asked for
  6. When a client leaves
  7. Plans, in a few lines
  8. Who it is for and who it is not for

Simulated scenario. Estudio Mirador is a fictional eight-person agency with twelve clients: an online shop, two clinics, an agricultural cooperative, a gym chain and seven more. The people, the agency and the numbers are made up and do not represent real customers or results. What we describe about the product, on the other hand, has been checked against our technical architecture, limits included.

When the promise lives in an email from six weeks ago

It is Tuesday. Norte Clinic asks whether the monthly report will now include call data, as discussed. Carmen, who ran that account until October, is now on two others. Nobody else remembers whether it was "in November" or "when the new dashboard is ready". Someone searches the inbox, someone the client chat, someone the notes of a meeting that was recorded but never transcribed.

In an agency with twelve clients, the most valuable knowledge is the hardest to find: what was promised, to whom, by when, and which exceptions were agreed. And the second problem is worse than the first: if you put all that memory in one place so you can search it, you need to be sure client A's material never shows up in client B's work.

A shelf of colour-coded archive boxes, one of them pulled out slightly

What would change with and without memory, with illustrative numbers

Item (per quarter, invented assumptions)BeforeAfter
"What did we tell the client" searches per week2525
Average minutes per search123
Forgotten promises that end in a complaint41
Handover meeting when an account owner changes1 hour20 minutes

With those assumptions, 25 weekly searches dropping from 12 to 3 minutes is about 3.75 team-hours a week. Do not trust our figure: it is a starting estimate. What is worth measuring in your agency is how many times a month someone asks "what did we tell...?".

How what was promised gets recorded

The rule in the scenario is simple: every promise to a client is recorded at the moment it is made, as an event in that client's project. One event per promise, with a type (decision or customer contact), a one-line title, the why in the body and an importance of 3 if money or a deadline is committed.

When Carmen ends a call, her agent (say Claude Code or ChatGPT connected over MCP) records: "Norte Clinic: the report will include call data from January, not earlier." When the deadline changes, a new event is recorded that supersedes the old one; the old one stops feeding answers but stays in the history.

Six weeks later, whoever needs it asks with mf_query ("what did we promise Norte Clinic about call data?") and gets the answer with its citation: the event, the date and who recorded it. And when the account owner changes, the morning brief and the project summary (mf_recall) do half of the handover.

A project manager on the phone looking at a wall calendar

Separation between clients: what exists and what does not

Precision matters here, because this is where agencies put trust at stake. What MemoryFirst does today, according to our architecture documentation:

  1. A workspace is a tenant. Every query is written with a tenant filter. As a second barrier, the database can apply row-level security, which switches on only if a self-test at boot passes; its actual state can be read from the service itself.
  2. A workspace holds projects. Each client can be a project. That is an organisational and retrieval separation (questions scoped with project only look at that client), but an unrestricted credential sees every project in the workspace.
  3. Credentials can be narrowed. A per-app grant limits an API key or an OAuth connection to certain projects, in read or write mode, with an expiry. A grant only narrows, never widens. In the API there is also per-project access for each person: anyone who is not a founder only sees the projects they are granted.
  4. A person can belong to several workspaces. The MCP server URL and its tenant parameter choose which one they act on, and an OAuth connection is tied to the workspace chosen on the consent screen.

What does NOT exist, and we will not hint otherwise:

  • There is no database per client. It is in our plan, not in production.
  • There are no per-client encryption keys on by default. Per-tenant encryption is implemented but off unless configured.
  • The second wall has a limit stated in the code itself: the connection is still the application's, so a SQL injection that managed to switch roles would lift that second barrier. We do not claim that leaked database credentials cannot cross between clients.

What to choose depending on what you have been asked for

  • Ordinary client, no special clauses: one workspace for the agency, one project per client, and narrowed credentials for apps and outside collaborators. This is the convenient option.
  • Client with a contractual demand for separation (a clinic, a law firm): a workspace of their own. The boundary is the tenant, not a project. That is the strongest separation we offer today.
  • How many workspaces each plan allows and how they are billed: we have not documented that here. Ask us before designing twelve workspaces; we do not want you to assume it from this article.

One more caution: if your client is in healthcare or law, do not promise outcomes or compliance because of a tool. The database, embeddings and files are in the EU; by default, the question and the retrieved passages go to a US inference provider to write the answer, unless you bring your own key (it covers mf_query and state regeneration, not every process). And there is no self-hosted edition.

When a client leaves

The exit matters as much as the start. A workspace owner can export it and request its deletion with a grace period, and there is a verifiable erasure that deletes derived data, checks nothing remains and issues a certificate. Export and account deletion are recent features: check them in the data settings before putting them in a contract.

One person hands a closed folder to another at the end of an engagement

Plans, in a few lines

Free: one project, 30-day retention, one seat. Pro: €12 a month billed annually (€14.40 month to month), unlimited projects and retention, one seat. Team: €39 a month billed annually, five seats and shared memory; each seat beyond five costs €9 a month billed annually (€10.80 month to month). VAT included, figures from our pricing page as of 2026-11-14.

Who it is for and who it is not for

It is for small agencies and consultancies where each account's knowledge lives in heads, emails and chats. It is not a CRM or a client portal: it is not meant for your agency's clients to query the memory directly.

Keep reading with your company's knowledge is scattered, what an AI company brain is and memory you can audit. To compare other options, see the AI memory market in 2026.

Keep each client's memory apart

Start free with one project and see how it behaves with a single client's data before you scale up.

Start free

Keep reading