Search returns hits, not answers
Full-text search finds ten documents. Whoever needs the answer reads all ten. For questions that span the whole body of knowledge it does not help at all.
Knowledge · Live
scroll Corporate Memory
Company knowledge does not become searchable. It becomes answerable.
The information has been in the building all along, the answer still is not. Corporate Memory brings your sources together into one body of knowledge and answers questions from it. Every statement carries its source, and anything unsourced is marked as such.
Where it breaks down
The information exists. The answer does not.
A hundred data sheets each state a load limit. The question of which product carries 25 kilonewtons is still unanswerable, because no single document compares them all. That is not a data problem, it is an access problem, and it repeats itself in every company.
Plausible wrong numbers
A semantic search cannot add up. Ask it for a number and it answers anyway. The answer sounds right and is not, and that is precisely what nobody notices.
Knowledge leaves with people
What experienced colleagues carry in their heads is written down nowhere. It walks out with a resignation or a retirement, and nobody knows in advance what was missing.
The solution
The data does not decide the technology. The question decides.
Questions from everyday working life fall into a handful of kinds, and each needs its own mechanic. Corporate Memory keeps all of them ready and picks the right one per question.
Every source passes through the same door into one shared store. What grows out of it (full text, comparison tables, relationships, timelines, condensations) are preparations of that single body of knowledge. They can be rebuilt at any time, switched on individually and switched off again.
The decisive difference is with numbers. Comparing, counting and sorting is its own mechanic, not a matter of phrasing. That is why one document type becomes a real table, and every evaluation over it carries a line stating how complete the underlying data actually was.
Sources where a copy from last night would not be outdated but simply wrong (task lists, calendars, stock levels) are not copied. They are queried at question time, using the sign-in of the person asking.
Not everything has to move for that. Where knowledge already sits in structure (in the ERP, in document management, in the ticket system, on the intranet), Corporate Memory docks onto the existing system instead of building a second truth beside it. Only what a preparation needs gets copied. The rest stays where it is and is retrieved at the moment of the question. That is the difference between another data island and genuinely connecting what is already there.
Look it up in the textCompare numbersFollow relationshipsChange over timeCondensed assessmentQuery live
Three principles
Soften any one of them and you have built a different product.
These three points are the reason Corporate Memory is built the way it is. They do not live in a manual, they live in the data model and in the code.
Intake contract
Nothing enters the store without origin, time, confidentiality level and validity. That is not a step somebody can forget, it is a condition of the database. A file uploaded in the chat goes through the same door.
Source obligation
Nothing may look sourced that no source backed. An answer without a named citation is marked as such and carries neither document nor page nor passage. „I do not know" is a correct answer. A plausible wrong number is not.
Permissions in the data layer
The usual solution checks permissions in the application. That holds until the first other tool reaches the database directly. Here the database itself checks, so every new surface is automatically as safe as the first.
In practice
Three questions, three mechanics, three sourced answers.
The same surface, but behind it something different happens depending on the question.
What does the contract say about notice periods?
The answer arrives with document, page number and the passage it came from. Anyone who wants to read it is one click away.
Which product carries 25 kilonewtons?
A comparable table was built from the data sheets beforehand. The answer sorts and compares across it and states how many data sheets contained the value at all.
What is on my task list today?
This question goes live to the source system, using the sign-in of the person asking. Two people get two answers, and it says so.
Dedicated permission management, in all three cases
The same check precedes every one of these answers: what is the person asking allowed to see at all? The knowledge sits centrally and completely, and is still visible in tiers. How that works is in the next section.
Permissions and visibility
Everything in one place. But not everything for everyone.
A central body of knowledge and tiered visibility are not a contradiction. They are the condition under which such a system may be introduced at all. Every document carries a confidentiality level, every person a clearance level, and access rights can additionally be inherited from the source system.
- Open within the companyManuals, standards, product documentation, process descriptions. Visible to everyone who is signed in. The bulk of the knowledge sits here.
- ConfidentialContracts, costings, project files. Visible to the departments that work with them. Anyone without clearance does not even see these citations listed as hits.
- Strictly confidentialPersonnel matters, negotiations, board papers. Cleared by name. Where even the title gives something away, the source reports only that there is a hit, not which one.
This is checked in the database, not in the application. Without a person context set, a connection sees nothing at all, and that is the default. The data catalogue is fully prepared regardless: a document somebody may not see is still captured, indexed and prepared. It is simply not visible to that person. That keeps the whole body of knowledge evaluable without anyone seeing more than they may.
How it works
Four steps from the file share to a sourced answer.
Not a project spanning months. The first source stands within a day, everything else grows from it.
Connect a source
An assistant guides you through five steps, from origin to review. File shares, interfaces and feeds all take the same route.
Choose preparations
Looking things up is mandatory, everything else is chosen by the question it should answer. Cost and duration are shown at the switch, before money is spent.
Ask questions
In the chat, with selected sources. The mode is stated before the question and again on the answer, so it never has to be guessed.
Keep it current
A schedule per source, anomalies when volumes drop, and open questions drawn from what the store could not answer.
Inbound: dock onto other systems
Corporate Memory can bind existing APIs and third-party MCP servers as a source and query them at question time, using the sign-in of the person asking. That makes systems answerable that we deliberately do not want to copy, such as task lists, calendars or stock levels.
Outbound: its own API and MCP server
The other way round, Corporate Memory exposes its own tools as an API and as an MCP server. Other applications and agents work on the same body of knowledge under exactly the same permissions as the person they act for.
Let us take one of your file shares and see which questions would already be answerable today.
Arrange a conversation Core capabilities
What the module brings.
Look it up in the text
Full text and semantic search combined into one result list. With real page numbers for PDFs.
Documents become a table
One document type turns into comparable rows. Every value must appear in the passage it cites, otherwise it is discarded.
Follow relationships
Which customers depend on a supplier, which parts on a standard. Across several levels as well.
Change over time
Dated statements make a timeline. What applied when, and since when something else applies.
Source obligation with marking
Every statement from the store carries its citation. Without one, the answer is visibly marked as unsourced.
Permissions in the database
Visibility depends on clearance level and inherited access rights. Without a context, a connection sees nothing at all.
Freshness on a schedule
Daily, weekly or manual per source. The state is visible, and overdue sources stand out.
Anomalies
A feed suddenly delivers twelve entries instead of five hundred, a field stops being filled. Both report themselves as a finished sentence.
Open questions
Whatever the store could not answer is collected and grouped, along with a suggestion of which source would close the gap.
Feedback with four reasons
Wrong, incomplete, nothing found, wrong source. Each reason points at a different construction site instead of a thumbs up or down.
Deletion with proof
One command removes a document or a whole source completely, including the condensations that rested on it. The log records what was deleted, when and by whom.
Access for agents
Other tools can use the body of knowledge under exactly the same permissions as the person they work for.
Where it works
Six areas where the answer is missing today.
Sales
Product comparisons across the entire catalogue instead of the three documents that happen to be at hand. Ready to quote during the conversation rather than the next day, and with a statement that is sourced.
Product development
What has already been tested, rejected and why. Old test reports, complaints and standard revisions become comparable instead of sitting on network drives.
Business development and project management
Earlier projects, quotes and lessons learned become retrievable. Costing assumptions and timelines can be checked against what actually happened.
Human resources
Works agreements, policies and processes answer themselves instead of being asked three times a week. Confidential material stays protected through the clearance levels.
Legal and procurement
Deadlines, clauses and terms with a citation. What is not in the store is marked as unsourced rather than invented.
Management
The experience of departing specialists stays retrievable. Open questions show continuously where the body of knowledge still has gaps.
Surfaces
The knowledge base is one. The way in is yours.
How people work with it is not fixed. The same answering logic appears where the question actually arises, and that is rarely another portal you have to open first.
- Chat in the browserThe full surface with source selection, citations and feedback.
- Inside the business applicationAs a context panel directly in the ERP, the ticket system or the design environment.
- On the phoneAs assistance during the call: the answer is there while the customer is still asking.
- In the messengerFor the shop floor, field service and everyone without a desk.
- On the desktopAs a slim search bar above all applications, called up by keyboard shortcut.
- As help on the websiteOutward facing, limited to a released section of the knowledge base.
All routes use the same answering logic and the same permissions. There is no path that bypasses the source obligation or the clearance level, and that is exactly why the number of routes can be increased without risk.
A look inside
Two views that make the difference.
On the left what the person asking sees: an answer with its citation attached. On the right what the person responsible sees: the state of a source. The coloured dots show the status, the question mark explains what the line means in the result.
An answer and its source
The selected sources sit above the question, the citations below the answer.
- Sources: contracts, data sheets
- Answer rests on 3 citations
- Framework agreement 2024-118 · page 7
- One statement without a citation: marked „unsourced"
A source and its preparations
Per source you can see what is active, how well it was last reviewed and what is due next.
- Look it up in the text · active · reviewed 98 %
- Compare numbers · active · coverage 87 %
- Follow relationships · not active
- Schedule daily 04:00 · overdue by 2 days
Where it sits
The knowledge layer of the platform.
Corporate Memory sits in the Knowledge impact area and shares platform, sign-in and operation with the other modules. Whatever is sourced here can be used as context by the rest.
Culture
Enable people, do not replace them.
Agents
Automate processes, create room to work.
Knowledge
Secure experience and make it usable.
Content
Scale communication, on brand and on quality.
Modules around it
Local AI ServerOperation on your own premises
Agent-as-a-ServiceUses the store as context
Data Migration AssistantBrings legacy holdings in cleanly
One store, one truth. Vector index, relationships and timelines are preparations of it and can be rebuilt at any time. Anyone who makes the search index the original can never cleanly delete, correct or migrate again, and that decision is made on day one.
Operation and data sovereignty
You decide where your knowledge lives.
Corporate Memory is delivered software, not a service you hand your documents to. One installation per customer.
Cloud
Operated in European data centres, with monitoring and updates during live operation. The fastest route to the first sourced answer.
Hybrid
The application with us, the data with you, or the other way round. Sensible when individual holdings may not leave the building.
On-premise
Entirely in your own network, on the Local AI Server if you wish. Then neither the question nor the citation leaves your building.
One quote says more than many promises
“We spent two years talking about a knowledge base. What made the difference was not the search, it was that every answer states where it comes from. Only then did people start using it for things that matter.”
Head of Engineering, Mechanical engineering, around 400 employees
Pricing
One setup, one predictable operation.
Tiered by sources, size of the holdings and the preparations switched on. Setup covers connection, first preparations and acceptance against a question catalogue.
| Package | For whom | Scope * | Setup * | Operation * |
|---|---|---|---|---|
| Starter | First holdings, one department | up to 3 sources · up to 25,000 documents (around 12 m words) · 25 GB · 10 users · look-up in the text | from €7,900 | €249 / mo. |
| Business | Several departments in regular use | up to 10 sources · up to 250,000 documents · 150 GB · 50 users · all preparations · extraction with cost approval | from €17,500 | €549 / mo. |
| Pro | Group-wide, with agent access | unlimited sources · from 500 GB · live sources · agent access · clearance levels and inherited access rights | from €29,500 | €990 / mo. |
CorporateSeveral installations, a dedicated operating environment or particular requirements for evidence and certification: on request.
Storage
Included are 25 GB (Starter), 150 GB (Business) and from 500 GB (Pro). Measured on the original files, because those stay in place. Additional storage costs €3.50 per GB and month.
Not only documents
Images, drawings and scans count towards storage, their text recognition is billed per page (around 1 cent). A scanned archive therefore needs noticeably more lead time than the same volume of text files.
Additional usage
AI usage in question-and-answer operation is included in fair quotas. Extraction runs are measured on a sample, extrapolated and approved individually. Guide value from 3 cents per document depending on the number of fields.
Operation includes maintenance, updates, model care and a monthly quota for AI usage. Server and hosting costs are billed separately. Operation on your own premises on the Local AI Server is offered as a combination.
* Guide values. Scope, volumes and prices are adapted to your actual needs in the quotation, because the number of sources, the share of media and the depth of preparation drive the effort.
All prices are net, plus statutory VAT.
Frequent questions
What decision makers ask first.
What happens when nothing in the store answers a question?
Then the answer arrives visibly marked as unsourced, without document, without page, without passage. The question is additionally recorded as an open question, so it becomes clear where the store has a gap. That is precisely the difference from a system that would rather answer something.
So does everyone see everything?
No. Every document carries a confidentiality level, every person a clearance level, and access rights can additionally be inherited from the source system. This is not checked in the application but in the database. A tool with direct access therefore also sees only what its person may see.
How do existing file shares get in?
Through a guided assistant. File shares with text and PDF documents take the short route, interfaces and feeds are prepared with sample data and confirmed once. Every file passes through the same intake as any other source.
What does an extraction run cost?
That is measured on a sample beforehand and extrapolated. The figure is shown at the switch, and the run only starts after explicit approval. Runs are unique per document, so repeating one does not double anything.
How does deletion work, and how do you prove it?
One command removes a single document or a whole source. Because the preparations are derived from the same database, they disappear with it. Condensations that rested on the deleted document are removed as well. The deletion log records identifiers and counts, never a title and never a passage.
The first step
One file share. One week. Ten real questions.
We connect one of your existing file shares and put ten questions to it that come up in daily work. After that you can see in black and white which of them are answerable and where the store still has gaps.