A founder works alone in a bright studio with a laptop, a notebook and a coffee in the morning

Simulated use case: a solo founder with 11 projects and one shared memory

6 min read

Simulated scenario: a day for a founder juggling 11 projects if she and her agents shared one memory with citations. Before, after and limits.

by Equipo MemoryFirst

Table of contents
  1. Before: the invisible cost of switching projects
  2. How it is set up
  3. A day with shared memory
  4. Before and after, with illustrative numbers
  5. What this scenario does not fix
  6. Next step
  7. Sources

Simulated scenario. Marta Olmedo, the person in this article, does not exist, and neither do her companies or her clients. It is an invented case to show how a shared memory would work on one concrete working day. The numbers are plausible but illustrative: they are not results measured at any real customer.

Marta founded a small software studio on her own and now keeps 11 projects going at once: two subscription products of her own, a mobile app, four websites for clients, an internal invoicing tool, a technical blog, a browser extension and an experiment that is still alive out of stubbornness. She runs them with Claude Code, ChatGPT for writing and an outside helper a few hours a week.

Her problem is not technical. Every switch between projects forces her to rebuild from memory where she was. It is the one-person version of your company's knowledge is scattered: what she decided on Tuesday is in a chat, what she agreed with a client is in an email, and her agent sees none of it, because it starts cold every session.

A wall of sticky notes and coloured cards arranged in columns above a desk

Before: the invisible cost of switching projects

On an ordinary Tuesday, before shared memory, Marta's day looks like this:

  • She opens the mobile app project and loses a good while rereading the last thread to remember why one payment gateway was ruled out.
  • A client writes asking about a deadline. She thinks she remembers a date, but is not sure whether she agreed it by phone or by email.
  • She asks her agent to carry on with a refactor and has to explain the project rules to it again, once more.
  • In the afternoon she finds that two projects use different prices for the same thing, because one was never updated.

Marta estimates, roughly, that she loses about 90 minutes a day rebuilding context. That number is invented for this scenario: what matters is the shape of the problem, which grows with every project added.

How it is set up

Marta is on the Pro plan (12 EUR a month billed annually, per the pricing page on 14 November 2026), because she needs unlimited projects and connectors. Each project is a separate memory space inside her account, so the mobile app's context does not mix with the clients'.

  1. Her Claude Code sessions feed the memory through the hooks: each session starts with the project's context and, when it ends, is distilled into decisions and blockers. The guide to memory for Claude Code describes the real procedure.
  2. Her email and calendar come in through the Gmail and Google Calendar connectors, read-only.
  3. Her contracts and quotes from a Google Drive folder are indexed with the page and fragment they came from.
  4. Call notes she writes by hand or pastes as text, because MemoryFirst does not transcribe audio: the text you bring is indexed as one more source.
  5. A morning brief on Telegram summarises what moved.

Without point 4 there is no magic: if she does not write the note down, the call does not exist for the memory.

A day with shared memory

7:40. The morning brief tells her a client replied about a website design and that a delivery is due this week (picked up from the calendar). She starts the day deciding, not reconstructing.

9:10. She opens Claude Code on the mobile app. The start hook injects the project's state: the payment gateway was ruled out because of a limitation documented in a session three weeks ago, and there is an open blocker with the store review. She does not have to explain it.

A woman walks down a sunlit street with a notebook while on a phone call

11:30. The client of one of the websites asks about the deadline. Marta queries the memory: "what deadline did we agree with this client?". The answer cites the email of 3 October and the later call note, which moved it a week. Both citations can be checked, so she replies without hesitating.

15:00. She switches to the internal invoicing tool and asks which price applies to the basic plan. With Pro's temporal graph, the memory spots two facts that do not add up: the website's price (older) and the one from a later decision. It keeps the most recent and marks the old one as superseded, with the source of each version. Marta fixes the website.

18:20. Before closing, she asks her agent to record a decision with mf_record: "Extension launch is postponed until the permissions policy is resolved." Tomorrow, any of her agents will see it.

A calm evening desk with a closed laptop, a paper planner and a cup of tea

Before and after, with illustrative numbers

These figures are invented for the scenario. Do not use them as an expectation of results.

SituationBeforeAfter (illustrative)
Daily time spent rebuilding contextAbout 90 minAbout 30 min
Looking up "what did we agree with this client"10 to 20 min, with doubtsAbout 2 min, with a citation
Re-explaining rules to an agent when opening a sessionEvery sessionOnly when they change
Contradictory data spottedBy chanceWhen queried
Cost of the tool0 EUR12 EUR/month (annual)

The drop from 90 to 30 minutes is not a promise: it is what would have to be measured in a real case, and we recommend timing yourself for two weeks before believing it.

What this scenario does not fix

  • It does not replace Marta. The memory recalls what someone recorded; if nobody notes a decision down, it does not exist.
  • Calls are not transcribed. You have to bring the text.
  • Inference is not EU-only. Data, embeddings and files are in the EU, but by default answers are written by a US provider (Cerebras), unless she configures her own key for queries. If a client demands otherwise, Marta needs to know before connecting their documents.
  • Isolation between the service's customers is by query filters plus an optional second barrier; a database per customer is a plan.
  • Keeping 11 projects going is still a lot. A good memory lowers the cost of switching context, not the cost of having too many contexts.

Next step

If you recognise yourself in Marta, start with one project, the one that hurts most, on the Free plan. To understand the product, read what a company brain is; for plans and price, see pricing; and to compare with other memory layers, how to choose a memory layer.

Sources

  • This is a simulated scenario: people, companies, projects and figures are fictional. There are no external sources to cite.
  • Capabilities, supported sources and data limits: internal product documentation checked against the main branch code on 2026-11-13. Prices: MemoryFirst pricing page (accessed 2026-11-14).

Try it with your projects

Start with one on the Free plan. If you work on several, Pro offers unlimited projects.

Start free

Keep reading