A work table covered in sticky notes, papers and open laptops, with several people searching through them

Your company's knowledge is scattered: chats, calls, documents and repos

Problem7 min read

Why what your company knows is split across chats, calls, documents and repos, what searching costs, and what it takes to bring it together with citations.

by Equipo MemoryFirst

Table of contents
  1. Where the knowledge really lives
  2. One question, five places: what searching costs
  3. Why searching each tool doesn't fix it
  4. What is lost when it isn't brought together
  5. What it takes to solve it
  6. How each option on the market solves it
  7. Sources

A client writes: "Didn't we agree the deadline was the end of the month?" To answer with confidence, you have to rebuild a story that isn't whole anywhere. Part of it was said on a call. Part is in a chat thread. Part is in the email with the quote, part in the PDF contract, and the last piece is in a commit message from two weeks ago.

This scene is made up, but any small team will recognise it. And it isn't a discipline problem: a company's knowledge is produced on many surfaces at once, and almost none of them store it in a form another can read.

Where the knowledge really lives

The table below shows a typical version. It is an illustrative scenario, not the result of a study.

SourceWhat it holdsWho sees it
Calls and meetingsDecisions, nuance, promisesWhoever was there (and, with luck, minutes)
Chats and channelsClarifications, changes of directionChannel members
EmailQuotes, agreements with clientsEach person's own inbox
Documents and contractsTerms, signed figuresWhoever has folder access
RepositoriesThe technical why, in commits and PR descriptionsThe engineering team

Each has its own search, its own permissions and its own way of ordering time. None knows about the others.

Notebooks, a phone, a printed contract and a laptop laid out separately on a desk seen from above

One question, five places: what searching costs

Illustrative: round assumptions, not measured. Imagine one person answers 3 questions a day of the "what did we agree with this client?" kind, and each takes 8 minutes of hopping between tools.

That is 24 minutes a day, about 8 hours a month (at 21 working days). In a team of six, nearly 50 hours a month spent remembering instead of deciding.

And that is the good version, where the answer exists and can be found. In the bad one, the person fills the gap with what they think they remember, and someone pays for the mistake later.

Why searching each tool doesn't fix it

The logical reaction is to improve each tool's search. It helps, but it doesn't solve three things.

  • The answer is split: the question "what was the deadline?" needs a sentence from the call, the chat confirmation and the contract clause all at once. A per-tool search returns loose pieces.
  • It doesn't know which is current: if the price changed between two emails, both show up and nobody knows which one rules.
  • It returns text, not reliable provenance: you can find a fragment, but you have to work out where it came from and who said it.

It gets worse when the one asking is an AI agent. A person is suspicious of a doubtful answer; an agent uses it. We cover that other side in your AI agents start cold every session.

What is lost when it isn't brought together

The hours are what you see. The invisible part is usually what costs most.

  • The why behind decisions: the document says what was chosen; the conversation says why the alternative was dropped. Without the why, closed debates get reopened.
  • Contradictions: a fact changes in one place and stays old in others.
  • Knowledge that leaves with people: whoever left the company took half a client's history with them.
A person comparing two printed sheets of paper side by side with a puzzled look

What it takes to solve it

For the client's question to be answered in a minute and with confidence, a solution needs six things.

  1. Take in every surface, without changing anyone's habits: conversations, documents, email and operations, not just PDFs.
  2. Structure what comes in: facts, entities and dates, not just chunks of text.
  3. Keep it current: when two facts clash, keep the most recent and mark the earlier one as superseded.
  4. Cite the exact origin: the page, the minute or the message, so every answer can be checked.
  5. Be readable by any AI, not just one, so changing model doesn't force a migration of your knowledge.
  6. State precisely where the data goes, with no round-number claims that aren't true.

What follows is what MemoryFirst does today on each point, and where the limit is.

Which sources you can bring today

Working today: the events API, file upload, Gmail and IMAP, Google Calendar, Drive, Dropbox, OneDrive, Slack and Telegram, plus Claude Code sessions. Voice means transcripts you bring yourself; there is no audio channel. WhatsApp Business, Linear and Sentry are planned, and there is no automatic git repository connector. We say it this way so nobody buys something that isn't there.

How currency and citation work

Each fragment carries a reference to its origin (page, minute range, message identifier or URL). When two facts contradict each other, the graph keeps the most recent sourced one. We explain it in how MemoryFirst retrieves and in memory you can audit.

Where the data goes

Database, embeddings and original files are in the EU. But by default the question and the retrieved passages go to Cerebras (US) to write the answer, unless you bring your own model key. If your sector requires EU-only inference, we can't offer it today.

A small, relaxed team in a bright meeting room around a table with one laptop

How each option on the market solves it

We summarise what their public pages say, accessed 2026-11-14. Where we couldn't check something, we say so.

  • A wiki or document manager: works when someone writes. Knowledge from calls and chats rarely makes it in, and it goes stale if nobody maintains it.
  • Glean: presents itself as an enterprise AI platform connected to a company's data, with search, an assistant and agents. It claims more than 275 connections to work apps. It shows no public pricing: it only offers a demo request. Its headline talks about "enterprise scale"; we could not verify from what customer size it is offered. More in MemoryFirst vs Glean.
  • Zep: a context layer for agents. Its Flex plan ($125/month) lists connectors to Google Drive, Slack, Notion, Gmail and Granola. Aimed at people building agents over an API. More in MemoryFirst vs Zep and the comparison page.
  • Mem0: a memory layer for agents and applications, over an SDK or API. On the pricing page we consulted we did not see connectors to company sources described. More in MemoryFirst vs Mem0 and the comparison page.
  • Supermemory: the Gmail connector appears from its Max plan ($100/month), and the S3 and web crawler ones in Scale ($399/month). See the comparison page.
  • MemoryFirst: connectors for email, calendar, documents, Slack and Telegram, per-fragment citations and currency in the graph, reachable over MCP from any AI client. Limits: voice means transcripts, no audio; default inference is not EU-only; and there is no self-hosted edition. The Free plan includes one project and 30-day retention; the others are on the pricing page.

For a definition of the missing piece, read what a company brain for AI is. If you work for several clients at once, see the use case of an agency with many clients, a simulated scenario.

Sources

Third-party data, accessed 2026-11-14:

  • Glean, home page: https://www.glean.com/ (accessed 2026-11-14)
  • Zep, pricing: https://www.getzep.com/pricing (accessed 2026-11-14)
  • Mem0, pricing: https://mem0.ai/pricing (accessed 2026-11-14)
  • Mem0, home page: https://mem0.ai/ (accessed 2026-11-14)
  • Supermemory, pricing: https://supermemory.ai/pricing (accessed 2026-11-14)

The search-time figures are illustrative assumptions, not measurements.

Bring together what your company knows

Connect your sources and query with citations. Start free with one project, with data stored in the EU.

Start free

Keep reading