A small founding team gathered around a wooden table with laptops, notebooks and sticky notes on the window behind them

Why MemoryFirst was born: a company's memory isn't where you look for it

6 min read

The short story of MemoryFirst: the problem we saw, why we built a memory layer with verifiable citations, and why we decided to build it ourselves, in Spain.

by Equipo MemoryFirst

Table of contents
  1. The problem: real knowledge isn't in the CRM
  2. Why what existed wasn't enough
  3. Why build it now
  4. The thesis: memory is the part you can't swap
  5. Why us
  6. What we decided not to do
  7. What is true today and what is a plan
  8. Who it is for and who it isn't

Picture the scene. On Tuesday, a plan change gets decided in a twenty-minute call. Nobody records it, nobody writes minutes, and two people drop a loose sentence in a chat. Three months later a client asks, "Didn't we agree on a different deadline?" The answer they get is a mix of memory, searches across three tools and a guess delivered with confidence.

That scene, repeated in every small company, is why MemoryFirst exists. This is the short story: what we saw, why we decided to build it, and why we are the ones building it.

The problem: real knowledge isn't in the CRM

A company's knowledge lives in last Tuesday's call, in the thread between two colleagues, in the contract signed three months ago and in the meeting where the pivot was decided. Tools that index PDFs capture a fraction of that, and the rest gets guessed.

What changed in the last few years is who asks the questions. It used to be a person digging through a folder. Now it is also an AI agent, which has no intuition to fill in what is missing: it either gets the right fact or papers over the gap with something plausible.

A person taking handwritten notes during a call while a colleague types in a chat in the background

Why what existed wasn't enough

We saw two different gaps that meet at the same point.

The first is agents that start every session with no past, which we cover in detail in your AI agents start cold every session. The second is teams whose knowledge is scattered across chats, calls, documents and repositories, which we develop in your company's knowledge is scattered.

The available solutions covered part of it. Some index documents and leave conversations out. Others remember a user's preferences inside one product, but not a company's decisions. Almost none returned the answer with its exact provenance (the page, the minute, the message), which is what lets you trust it.

Why build it now

There were three reasons, and all three can be checked.

  1. The market is asking for a "company brain": memory that is structured, current and usable by agents. We lay that out in our welcome post.
  2. Citation quality is now good enough for business: with 2026 models, and careful retrieval, citing the exact source works.
  3. MCP makes the memory layer portable: build it once and plug it into Claude, ChatGPT, Cursor or your own agent. See how in plug memory into any agent.

The thesis: memory is the part you can't swap

Models are replaceable. A company's curated, verifiable memory is not. If there is lock-in, it lives in the memory layer rather than the model layer, which is why portability matters to us as much as retrieval quality: export, verifiable erasure and bringing your own model key (BYOK).

That thesis has an uncomfortable consequence for us. If the memory is yours, you have to be able to take it with you and check what we store. That is why every answer carries its citation and erasure leaves a certificate. We go deeper in memory you can audit.

An open wooden archive cabinet with neatly organised folders, a picture of well-kept memory

Why us

We build MemoryFirst with MemoryFirst. It is the tool we work with every day, and a single founder runs ten SaaS products on this brain (we say so on the home page). AlmaOS, a no-code product for firms and clinics that reasons across calls, emails, contracts and minutes, is built on the same engine.

That gives us something you can't buy: we know what breaks when a memory contradicts itself, when a citation lands wrong, or when a morning briefing is noise. And it is built in Spain, with the database, embeddings and original files hosted in the EU.

What we decided not to do

A product is defined as much by what it refuses as by what it includes.

  • It is not a notes app: people write elsewhere, the AI reads here.
  • It is not a CRM or a chatbot builder: your customers query their memory, they don't build bots.
  • It is not an agent runtime: memory and retrieval only.
  • It does not train models on customer data, and we will not build a foundation model of our own.
  • We don't build a messaging app: we integrate the ones you already use.
  • We don't promise isolation, residency or encryption that the code and infrastructure don't yet provide.

What is true today and what is a plan

We would rather say less and have it be true. This table separates what works today from what is still to be built.

CapabilityStatus
Ingestion of events, files, Gmail/IMAP, Calendar, Drive, Dropbox, OneDrive, Slack and TelegramIn production
Verifiable citations (page, minutes, message, URL)In production
Remote MCP with OAuth 2.1 or API key, 15 toolsIn production
Verifiable erasure with a certificateIn production
Call audio, Linear, Sentry, WhatsApp BusinessPlanned (voice means transcripts you bring yourself)
Database per customer, per-customer encryption on by defaultPlanned
EU-pinned inference per customerPlanned
Self-hosted edition of the productDoes not exist

On data, the exact sentence is this: your database, embeddings and original files live in the EU (dedicated infrastructure in Germany). But by default the question and the retrieved passages are sent to Cerebras (US) to write the answer, unless you bring your own model key. We don't say "everything in the EU" because it wouldn't be true.

A data centre corridor between server racks lit in cool blue light

Who it is for and who it isn't

MemoryFirst fits three profiles: small teams and firms (clinics, accountancies, consultancies) that lose hours hunting for "what did we promise this client"; teams working with AI that don't want to build a memory layer of their own; and anyone who needs to know precisely where their data goes.

It is not for you if you are after a notes app, a CRM, a chatbot builder or a place to run agents. Nor if you need EU-only inference or a self-hosted edition today: neither exists today, and we would rather tell you before you find out.

If you want to understand the central piece, read what a company brain for AI is. If you would rather try it, the Free plan includes one project, MCP memory with the Claude Code hooks and 30-day retention; the other plans are on the pricing page. And to see how it compares with other options, start with the AI memory market in 2026.

Give your AI a memory

Start free with one project: MCP memory, Claude Code hooks and data stored in the EU.

Start free

Keep reading