FAQInsightsFRENBook a call
Research

Research

Company memory is the subject that conditions every AI use to come. An agent that does not know the company can neither reason correctly nor act at the right moment.

We made it our specialty.

Measured on the public reference benchmark, our company memory delivers today the best published answers, three times faster and twice as cheap as the best alternative. We devote continuous research to it. What we deploy at our clients comes from it, tested, retested and put to the proof in real company settings.

92.3

out of 100 in answer quality, the best published score on the reference benchmark.

3×

faster than the best competing system: 21 seconds for a complex question across 500,000 documents, versus 74, and 90 for an agent that opens files one by one, with no company memory.

2×

cheaper per question than the best competing system: $0.47 versus $0.84, and $1.35 for an agent that opens files one by one, with no company memory.

01

What we master

We build company memories: all of an organization's information, linked and queryable, with the same level of context as an employee who knows the company. The tools in place are kept, we connect them. This work raises six questions that few players have settled. Here is how we answer them.

The real risk with enterprise AI is not that it makes up an answer, it is that it misses information that exists. A classic search only returns the documents that resemble the question; when the answer is spread across hundreds of documents, it misses half of them. Our memory links each piece of information to the ones that concern it, and retrieves 99.5% of the documents needed to answer.

Complete answers and a cost per question divided by two. Every person, every company, every meeting exists once in the memory, and the relations between them are established from verifiable information, never guessed.

Every piece of information and every relation keeps its source. An answer always says where it comes from, when, and from whom. When two sources contradict each other, the memory flags it instead of choosing for you.

Access rights are attached to the information itself, not to the tool that reads it: a document only management can read is only accessible to management, whether it is searched from Claude, from ChatGPT or through an agent. And those rights apply to agents exactly as they do to people.

An agent that reads the memory cannot send anything outside your company: it has no access to the Internet or to your messaging tools. And it is never the language model that decides what it is allowed to see or do: those rules are set in stone, outside of it, and it cannot bypass them.

The model that reads the memory is an interchangeable part. You switch it without rebuilding anything, and you use the memory from the tool of your choice.

02

What we measured

Our company memory against the public reference benchmark

We measured our company memory on EnterpriseRAG-Bench, a public, open benchmark. It reproduces the data of a fictional company, more than 500,000 documents from chat, email, tickets, shared documents and a CRM, and asks 500 questions whose reference answer is known. We ran it with exactly the system we deploy at our clients, on 42,583 documents. The times and costs below are measured on that corpus, far larger than a conversation with a consumer assistant.

Our company memoryThe best competing systemAn agent opening files one by one, with no company memory
Answer quality, score out of 10092.390.686.0
Time per complex question21 s74 s90 s
Cost per complex question$0.47$0.84$1.35

More accurate answers, three times faster and twice as cheap as the best alternative. The other systems' scores come from their own publications; the time and cost gaps are unambiguous.

What our architecture brings

We measured what the links between documents bring. On the same corpus of 42,583 documents, we asked 47 questions of the kind "which employees exchanged with this client?", whose answers require going through 1 to more than 300 documents. Classic search is what AI assistants plugged into your documents do, like ChatGPT or Copilot: they keep the documents that most resemble the question. Same system, same model: the only difference is the network of links.

1 to 10 documents
100%
76.5%
10 to 30
100%
68%
30 to 100
99.4%
58%
more than 100
98.5%
50%
Overall
99.5%
60%

With the links, the system retrieves all the answers and costs 2.3 times less per question. It is better on 42 questions out of 47 and never worse, measured three times in a row. Without the links, even on simple questions, a quarter of the answers are missing. The more history your company has, the wider the gap.

Concretely

When your company has years of history, the graph is what guarantees that no answer is missing.

03

How a company memory works

What we deploy at our clients comes from our research, tested, retested and put to the proof in company settings. Here are the principles that make it reliable.

One source of truth, readable by a human

All information converges into a single repository, written in a format anyone can open and read. The index that serves searches is derived from it and can be rebuilt at any time. Nothing is locked in a database only a vendor knows how to read.

Collect continuously, consolidate permanently

Ingestion agents capture what is new in tools, meetings and exchanges, structure it and file it. The system then rereads the whole, links what should be linked, spots what has changed and what contradicts itself. The memory gets more accurate every day.

Link rather than pile up

A person, a company, a meeting, a decision exist once. The relations between them are derived from structured, verifiable information, never guessed from text. Every relation keeps its provenance: it is always possible to go back to the source that states it.

Everyone sees what they have access to

Like a shared file space, each user opens only what they are allowed to. Confidentiality is decided where the information lives, not in the tool that reads it, and it applies to agents as well. An executive, a case manager and an agent do not access the same content.

Agents that cannot compromise anything

An agent that reads the memory cannot send anything outside your company, and it is never the language model that decides what it is allowed to see or do: those rules are set outside of it. Everything that enters the memory is validated before it stays, because a false piece of information that spreads is as serious as a leak.

Model-independent

The memory belongs to you and lives on your side. It is used from the tool of your choice, and the model that reads it is an interchangeable part: you switch it without rebuilding anything.

Every agent leaves a trace

What it read, what it did, why. That is what makes it possible to detect a regression, correct and improve over time.

And in your company, what would it look like?

In 30 minutes, we show you the Company Brain in a demo, and how it would connect to your existing tools, replacing nothing. You leave with a clear idea of what it would change for your teams.

Book a callSee the Company Brain in productionContact us by email