Company brain · EU data residency

Your AI forgets everything. Give it a memory.

Your editor's memory plugins remember your coding style. MemoryFirst remembers your business: decisions, clients, deadlines, with verifiable sources and data resident in the EU. Claude Code, Cursor or ChatGPT start every session knowing where you left off, so you stop re-explaining your own business to a machine.

Start for free Founder Lifetime: 9 EUR/mo forever

14-day trial · No card required · Open MCP · EU data · We never use your data to train models

claude · no memoryWithout MemoryFirst
> continue the Acme onboarding
Where were we? I don't have context from
previous sessions. What is Acme and what
had you decided so far? 

claude · with MemoryFirstWith MemoryFirst
> continue the Acme onboarding
mf_recall(project: "acme")
Picking up where you left off: Acme is moving
to the Team plan (decided Tuesday); the blocker
was SSO. Next step: verify the domain. 
Source: decision 2026-07-15 · #sales (Slack)

Ingest · four source families

One memory · EU-resident

Serve · every AI client

Hetzner EPYC · Falkenstein, DE · EU-resident

Example output →AUDIOconsult-2026-03-14.mp3 · 14:22–14:44PDFmsa-acme-2025.pdf · p.12CHAT#front-desk · C091F3K2

The proof

One solo founder runs 10 SaaS products on this brain.

MemoryFirst is not a prototype: it's the tool MemoryFirst itself is built with, every day. This is what a company brain does for its founder.

Morning brief

telegram · 07:30

Good morning. 3 things moved overnight:

Meridian Labs payment failed (2nd retry today)

API v2.4 deploy is green

Acme replied: SSO unblocked

It knows what happened while you slept

Every morning, a Telegram message with what moved in your company: payments, deploys, clients, deadlines. You start the day deciding, not reconstructing.

Contradiction resolved

graph · contradiction

Pro plan: 24 EUR/mo (email, May 12)superseded

Pro plan: 29 EUR/mo (decision, Jul 8)current

Conflict resolved: the most recent sourced fact wins.

It notices when two facts don't add up

If the price on your site contradicts the one in May's email, the brain flags it, keeps the most recent fact and marks the old one as superseded. With the source of each version.

Compiled skill

skillify · early access

Procedure repeated 3 times: publish release

Compiled as skill: /release

7 steps · verification checklist included

It turns repetition into something executable

Ask for the same thing three times and the brain compiles it into a skill: a procedure your AI runs end to end, with the steps and checks it already learned from you.

An honest comparison

Code memory and business memory are not the same thing.

Memory plugins for your editor are good at what they do; we use one daily ourselves. They do not compete with MemoryFirst: they remember different things.

Memory plugins for your editor

claude-mem, Mem0 and friends

They remember preferences, patterns and style of your code across sessions
They live on your machine or in a US cloud
Built for developers: installed and configured in the editor
Great so your agent never forgets how you code

MemoryFirst

The memory of your business

Remembers your company: decisions, clients, deadlines and the why behind them
Temporal graph with citations and resolved contradictions
Data resident in the EU, GDPR by default
A finished product: no SDK, no infrastructure to run

You can use both at once: one remembers how you code, the other remembers your company. See the full comparison

How it works

Your company brain, up and running in one morning.

Connect your AI in 2 minutes

One command installs the MCP server in Claude Code, Cursor or Claude Desktop. From that moment your AI remembers across sessions: decisions, context and the reasoning behind them. No migrations, no switching tools.

Connect your sources

Gmail, Calendar, Notion, Drive, Slack, WhatsApp, Telegram. The brain feeds itself in the background and organises what comes in into a graph of facts, each with its source. You change none of your habits.

Get your brief every morning

A daily message with what moved in your company while you weren't looking: payments, clients, deadlines, pending decisions. Like a chief of staff who never sleeps and always cites sources.

AlmaOS · the product on top of the memory

AlmaOS: this is what gets built on top of the memory.

AlmaOS is the partner that remembers everything: the no-code business product built on top of the MemoryFirst memory layer . It reasons across calls, emails, contracts and minutes, and answers pointing to the source of every data point. These are three of its territories.

Tax and payroll advisory

> “For client Talleres Ruiz S.L., what is deductible this quarter? Show me also the consultations where I flagged inspection risks.”

Cross-references invoices, bank statements, your advisory notes and call recordings. Multi-client view. Detects discrepancies between what was agreed on the call and what was billed. Never miss a deadline again.

Connected sources · Recordings · invoices · bank exports · regulations · client emails

Law firm

> “How many family cases did we win this year? Who is our strongest lawyer in contested divorces involving minors?”

Cross-analysis of case files, hearing minutes, client calls and emails. Compare performance by case type, success rate and average duration. Detect patterns you would never see reviewing case by case.

Connected sources · PDF case files · email threads · hearing recordings · client calls

Clinic (dental / medical / aesthetic)

> “How many of Dr. Pérez's procedures had post-operative complications? What did the patient say on the follow-up call?”

Cross-references clinical records, post-operative call transcripts and treatment notes. Identifies patterns that lead to complications. Reviews informed consent quality from recorded conversations.

Connected sources · Records · follow-up recordings · clinical notes · PDF consents · reminders

Looking for the finished product for your firm or your clinic, no code required? The partner that remembers everything, on this same engine.

Meet AlmaOS

Connect your stack and your AI clients

Your sources go in. Your agents read. No lock-in.

Native connectors for your email, documents, code and applications. On the AI side, it plugs into Claude, Cursor, ChatGPT, Claude Code and Codex over MCP, plus an OpenAI-compatible API for whatever you build on top.

▶For technical profiles

For technical teams: native OAuth connectors, OpenAI-compatible API and MCP-compatible server (stdio, HTTP and SSE transports).

Available now

NotionGoogle DriveGmailGoogle CalendarGitHubLinkedInClaude CodeCursorCodexClaude DesktopChatGPT

Coming soon

SlackWhatsAppTelegramOutlook + TeamsLinearJiraHubSpotSalesforceAttioDropboxZoomn8nZapierAudioMap (audio)AgenteUno (llamadas)

The engine, up close

Persistent, graph-based, verifiable memory. For the people building with AI.

Developers, technical founders and solo builders: stop re-explaining your stack and your business to every model, every session.

01Persistent contextshared, not per-session

Your agents and chats stop starting from zero.

Every new session you re-explain who you are and what your company does. The knowledge lives on one person's machine and never gets shared. MemoryFirst is a persistent, shared memory layer: your agents recall the relevant context on session start instead of asking you for it again.

▶For technical profiles

For technical teams: recall the project context at session start over MCP. No copy-pasting prompts, no private notes locked on a laptop.

// without a memory layer

session.start() · context = []

// with MemoryFirst

session.start()
context = mf.recall(project)

02Model-agnosticchassis and engine

Your knowledge is the chassis. The model is just the engine.

The memory is independent of the model and exposed over MCP. Swap Claude for Gemini, GPT or a local model without migrating anything. You do not marry a provider, and your context survives every model change. Multi-model and local-first.

▶For technical profiles

For technical teams: model-agnostic memory over a standard MCP server. The same knowledge graph feeds every client and every model you point at it.

ClaudeGPTGeminilocal

memory.layer · MCPmodel-agnostic

Graph over vectors

A graph of related facts, not a pile of similar chunks.

Vector RAG searches by semantic similarity and you have to pray it returns the right passage: it can quietly miss. MemoryFirst builds a graph of related facts (entities, not look-alike chunks). An invoice links to its service and its client. If the entity does not exist yet, it gets created. And the memory learns from the questions you ask.

Entities and relationships, not just top-k nearest neighbours.
Missing entity? It is created and linked, not dropped.
The graph grows as you query it: raw input becomes related knowledge.

▶For technical profiles

For technical teams: ingest to entity resolution to graph to answer with citations. Retrieval traverses relationships, so the answer does not depend on a similarity threshold landing on the right chunk.

vector RAG

top-k ~ 0.82 · 0.79 · 0.77
?

fact graph

invoice #2041 → service: onboarding → client: Acme

resolved entity · 3 edges

LLMs are not databases.

That is where the serious errors come from: ask a model to recall a fact and it will reconstruct something plausible, not retrieve the truth. MemoryFirst gives you verifiable memory with citations back to the source, so the answer is checkable instead of merely confident.

Every answer is verifiable

The source is part of the answer.

Click any citation and you go straight to the origin. The highlighted paragraph in the PDF, the exact minute in the recording, the specific message in the chat. No more 'trust what the AI said.' And if you need it for an audit, you have it.

AUDIOconsult-2026-03-14.mp3 · 14:22–14:44PDFmsa-acme-2025.pdf · p.12CHAT#arch-decisions · C091F3K2

every citation takes you to the exact fragment of the source

memory.queryok · 247ms

“What did we promise Mrs García about her March check-up?”

We promised at no additional cost the follow-up X-ray on 18 March, and that we would reassess the implant at that appointment.

AUDIOconsult-2026-03-14.mp3 · 14:22–14:44CHAT#recepcion · C091F3K2

What makes the engine different

A memory layer, not a search box with AI on top.

Four properties most 'RAG over PDFs' solutions skip.

01Sources5 source types

All your information, not just documents.

Recorded calls, conversations, documents, web pages, operational data. Documents are 10% of your real knowledge: we connect the other 90%. With native integrations for the tools you already use.

voice

124

docs

388

chat

2.1k

web

https://docs.acme.com/onboarding

ops

● deploy.staging2.1s

6.4k

one memory9,073 events indexed

02Verifiable citationsparagraph · minute · message

Every answer carries the exact source.

Takes you to the exact paragraph in the document, the precise minute of the recording, the specific message in the conversation. Every answer is one click from its origin. Real traceability, not just trust.

▶For technical profiles

For technical teams: page and position in documents · start/end in recordings · identifier in conversations.

AUDIOconsult-2026-03-14.mp3 · 14:22–14:44PDFmsa-acme-2025.pdf · p.12

CHAT#arch-decisions · C091F3K2

docpágina exacta

audiominuto exacto

chatmsg_id

cada respuesta lleva su fuente exacta, a un clic

03GDPR · EU dataEU · GDPR · data processing agreement

Full GDPR compliance. From day one, no extra setup.

Your data is processed on dedicated servers within the European Union. Data processing agreement included in all plans. We never use your information to train models. For businesses handling sensitive information, this is not a bonus: it is the minimum standard.

▶For technical profiles

For technical teams: self-hosted language models on dedicated EU infrastructure. Three deployment tiers available. DPA available before signing.

Region

EU · Falkenstein, DE

Datos en la UE · alojamiento dedicado · cumplimiento RGPD

EU SaaS

self-host

on-prem

04CompatibilityChatGPT · Claude · your AI

Works with any AI you already use.

Connects with ChatGPT, Claude and the AI tools you already use via native integrations and an open API. No lock-in, no changing your workflow. Scales with your team.

▶For technical profiles

For technical teams: MCP-compatible server (Claude Desktop, Cursor, VS Code) and OpenAI-compatible API.

Claude DesktopMCP

claude_desktop_config.json

CursorMCP

.cursor/mcp.json

ChatGPTopenai-api

GPT Action · base_url

Claude CodeMCP

.claude/mcp.json

Customsdk

mf.ts / mf.py

Verticals in production

What is already built on top of the engine.

Dental clinic · 12 people

The receptionist stops interrupting the doctor.

«I asked what we had promised Mrs García about her March check-up. It pulled the exact phrase from the consultation recording, minute 14:23, verbatim.»

Dental clinic · Valencia

Evidence

00:0014:22 → 14:4430:00

Dr. López: Dile a la Sra. García que reevaluaremos el implante en su visita del 18 de marzo. Le prometimos sin coste adicional para la radiografía de seguimiento.

AUDIOconsult-2026-03-14.mp3 · 14:22–14:44

Law firm · 28 lawyers

Every brief cites the contractual clause it relies on.

«Our associates draft four times faster because every relevant clause is one click away (page and PDF preview included).»

Boutique M&A firm · Madrid

Evidence

msa-acme-2025.pdf · p.12 · Apartado 12 - Residencia

página resaltada

PDFmsa-acme-2025.pdf · p.12

Startup · product team · 8 people

The whole team knows every client's context, no extra meetings.

«We connected contracts, call recordings and Slack. Now anyone on the team finds in seconds what was agreed with a client six months ago, with the exact citation. No more depending on someone's memory.»

B2B startup · Barcelona

Evidence

#arch-decisions · 2025-11-22 · 09:14

Júlia P.

¿Por qué separamos el gateway?

Residencia de datos: los clientes T2 self-host necesitaban anclar el LLM gateway en su propio k8s.

CHAT#arch-decisions · C091F3K2

Why memory is the moat

“Models become interchangeable; what is not is your business's structured, verifiable knowledge. MemoryFirst turns your sources into a graph of facts with citations, model-agnostic, that any AI can read: swap models without migrating your knowledge.”

Memory is the next defensible layer in the AI stack

Built in Spain. Data in the EU. Open where it matters (MCP).

Pricing