The Firmengedächtnis
What your company knows, in one place.
It collects emails, files, scans, photos and recordings, links them to cases and answers questions in plain language. Every answer shows where it comes from. And what only lives in people’s heads, it captures in guided interviews.
- Backed
- Every answer with its source
- In house
- Own instance, also without internet
- With your permissions
- Who may not see something, doesn’t
What did we promise Meyer in March about the delayed delivery?
A partial delivery by 28 March1 and a 3 % discount on the remaining quantity.2
| Item | Source |
|---|---|
| 1 | Phone note of 12/03/2026, section 2 |
| 2 | Email from S. Krüger to Meyer GmbH, 13/03/2026 |
Checked2 sources
Everything in one place, every answer with proof.
In most companies the knowledge is there, just scattered: in mailboxes, on drives, in folders and on phones. Search finds file names, but no connections. The Firmengedächtnis takes in every source, reads it, links it to cases and answers questions in plain language.
Sources
- Email (IMAP, Exchange, Microsoft 365, Gmail)
- File shares and Google Drive
- Scans and paper
- Photos
- Conversation recordings and audio
- Video
- Spreadsheets, exports and e-invoices (ZUGFeRD, XRechnung)
- Interviews
Firmengedächtnis
- Take inRecognise the format, extract the text
- UnderstandText and speech recognition, images
- LinkPeople, companies, cases
- FileWith origin and permissions
Use
- Questions in chat, with sources
- A case file with the whole history
- Your own AI tools via MCP
Finds what isn’t text
Scanned delivery notes, photos of the damage, conversation recordings. Text and speech recognition make them searchable.
A case, not a hit list
Request, quote, call and invoice belong together. You see them as one file, not as ten links.
No answer without proof
Every sentence points to the spot in the original. If nothing reliable turns up, it says so instead of guessing.
Permissions stay as they are
Whoever may not open a file on the drive won’t see anything from it in an answer either.
Secure knowledge before it leaves.
The trick at the machine, a customer’s quirks, whom to call when things get stuck: that is rarely written down. The Firmengedächtnis captures it in guided interviews.
- Department
Define the need
Says in its own words which knowledge is needed.
- Firmengedächtnis
Generate the guide
Questions with a weight (must, should, could) and a time estimate in minutes, plus a gap analysis with suggestions from the data.
- In conversation
Conduct the interview
One question after another. Pause and continue later.
- Employee
Review and release
Confirms every point. Only then is it usable.
Once or over several sessions, with one person or with several. A trial run, in which the system plays the participant, shows beforehand how the guide works.
For whom
Who is leaving soon. Retirement or resignation, often only a few weeks remain.
Who has already left. Former employees who are still reachable and want to help.
Who is the only one who knows. Experts on whom a topic depends.
Several experts
The same guide, several experts. The Firmengedächtnis puts their answers side by side, point by point. Nothing gets averaged: where two experts contradict each other is often the most important finding.
- Agreed
- Deviation
- Contradiction
Three ways to conduct the interview.
The Firmengedächtnis is Humexa compatible: speech technology and avatar come straight from Humexa. All three ways produce the same transcript, word for word. Only the person who spoke can release it.
Write
Type in the browser, on a PC or on a phone. For everyone who prefers writing.
Speak instead of type
Questions are read aloud, answers transcribed and can still be corrected. The speech technology comes from Humexa.
A conversation with a counterpart
Humexa’s realistically rendered avatar conducts the interview with voice and facial expression.
More in the box.
What comes with every instance, listed like the parts of a drawing.
| Item | Part | What it does |
|---|---|---|
| 1 | Questions by email | Send a question to the inbound address; the answer comes back with its sources, within your own permissions. |
| 2 | Source view | The page image with the passage marked, for audio with a time stamp. |
| 3 | Contradiction detection | Finds documents that contradict each other, for example an order confirmation against a phone call. |
| 4 | Cases with a timeline | Everything about a case in order of time, exportable as Markdown or PDF. |
| 5 | Knowledge spaces | Separate spaces to organise knowledge, for example by department or topic. |
| 6 | Roles and audit | Seven roles, an audit log and GDPR information requests. |
| 7 | Seven languages | Content in German, English, French, Dutch, Italian, Spanish and Polish. The answer comes in the language of the question. |
| 8 | MCP interface | Eight read-only MCP tools for your own AI tools. |
| 9 | File formats | More than 50 formats, from PDF and Office through email to audio, video and archives. |
Own instance, clear rules.
Company knowledge is sensitive. That is why every customer gets their own installation, and permissions come from the sources themselves.
Own installation
Every customer gets their own instance, in our data centre or on your own hardware. No shared databases, no shared models.
Also without internet
All parts run in an isolated network. The AI models run locally in house, including speech and avatar; external models can be switched off technically.
Permissions from the sources
Every text passage carries the permissions of its origin. What someone may not open on the drive never appears in their answers.
Traceable and deletable
For every finding it is recorded where it comes from. If a document is deleted, including in the source, it disappears from search and links immediately.
Fair in interviews
- The employee sees beforehand what is collected and who may read the result.
- Any question may be skipped: there are buttons for “I don’t know”, “I’d rather not say” and “Continue later”.
- Only the person who spoke releases their statements, nobody else.
- Statements stay statements: interview content is quoted as “Mr Weber said on …”, not as fact.
How we get started.
Small first, then wider. Four steps until the first interviews are running.
- 1
Pick a slice
One mailbox, one drive, one department. Start small.
- 2
First intake
Read in the data and check permissions with two real accounts.
- 3
Your own questions
Ask at least 50 real questions from daily work and evaluate how often the source is right.
- 4
First interviews
Start with the people whose knowledge would be lost first.
Let’s start with one slice.
In an intro call we work out which mailbox, drive or department makes the best start.
Developed and run by createIF Labs in Tuningen, the team that builds AI agents for production. Back to createIF Labs