04 · Library
A context library for the owner of a responsibility areaPeople get onboarded. Who onboards your AI?
Agents are appearing across the organisation, and each one gets its own instructions. Library brings the guidelines of your area into one library with an owner, versions and tests. Every agent reads the same current version, and you decide what changes.
The owner of a responsibility area such as People & Culture, quality, communications or expert work, with IT or the AI lead involved from the start
A library for your area, in your environment, used by all your agents
The library is built as a project, then kept current through maintenance
A few hours from the owner and experts, as interviews and short micro-sessions
Written guidelines, and a repository and runtime environment from IT. No agent platform needed.
Agents already read your guidelines. No one sees the whole picture.
The same rule lives in five places, five ways. Each agent builder handles the content differently: a PDF attachment, their own summary or another workshop with you.
The handbook gets updated. The agents don't. When a guideline changes, people get an announcement. Agents carry on with the old version.
AI reads text written for people. It doesn't know what "significant" or "without delay" means in your organisation, and it fills the gaps with general knowledge.
You hold the responsibility, but not the view. You answer for the content, but you can't see what agents tell employees and customers about it.
Three kinds of content, one library.
| Content | Examples | What is added for AI |
|---|---|---|
| Principles and rules | Values, employee handbook, quality policy | Interpretations, examples, limits and the points where AI must ask a person |
| Content and meanings | Brand guide, quality manual, project model | Terms, naming rules, diagrams as text and tone of voice |
| Expert judgment | One piece of expert work, such as proposal review | Decision rules, weightings, exceptions and where a person decides |
Every item has an owner, a version, a validity period and tests. The core content goes into the library and is developed to levels 3 and 4 of the handbook.
Change it once. Every agent updates.
Tell us about the change your way. An updated file, a written sentence or a voice note is enough. We prepare the change and test it.
See what changes before you approve. Old and new side by side, which agents use the content and what the tests show.
Approval is one click. The approval is recorded in your repository, and agents read the new version.
A one-page monthly report. What agents used, what they were asked without an answer, and which reviews are due.
Your experts talk. Our production system writes.
| Step | What happens | What we need from you |
|---|---|---|
| 1. Context Check and scope | We assess the current material and agree which core content goes into the library and which agents use it. | One meeting with the owner and IT |
| 2. Conversion and enrichment | The production system converts the material and asks experts for the missing interpretations in an interview and short micro-sessions, by voice if they prefer. | A few hours in total from experts |
| 3. Tests and approval | We test with a question set, with and without the library. The owner approves version 1.0. | The owner's approval |
| 4. Deployment | The library goes into your GitHub and is served to agents from your environment. | A repository and runtime environment from IT |
No workshops needed. When information is missing or contradictory, we ask a person instead of guessing.
The library always lives in your environment.
| You | Marram |
|---|---|
| The library repository, for example in GitHub | Gets narrow rights to prepare changes and tests, revoked in one step. |
| The MCP server agents use to read the library | Does not host the repository, the server or the logs. |
| Usage logs and their retention period | Sees aggregates only, without user identifiers. |
| Authority: nothing is published without the owner's approval | Does not run, publish or authorise agents. |
The library is open Markdown in Git. If you stop the service, the library and your agents keep working without Marram.
A project and maintenance. The library is yours.
The project is priced per responsibility area. We agree the price once the content, agents and acceptance criteria are scoped.
Maintenance is an ongoing service. It covers preparing changes, tests, the approval flow and the monthly report.
Expanding is straightforward. The next area uses the same structure, and agents can read several libraries.
Is this a company brain?
A company brain searches your organisation's sources. Library makes sure agents interpret your guidelines correctly: every guideline has an owner, a version and a test. The two work well together.
We already have agent management, such as Microsoft Agent 365. Do we need this?
Agent management handles agents' permissions and runtime. Library handles what agents know about your guidelines. The library connects to your agent management.
Do you build the agents?
Not in this service. Your team or partner builds them and connects them to the library. A workflow can also be delivered as Workflow Build.
Does the library replace the expert?
No. It also tells agents when to ask a person.
What happens if we stop?
The library, server and logs are already yours. Marram's rights are revoked in one step.
Do we need a particular AI tool?
No. The library is written in open formats that different AI tools and agents can read. You can switch tools without rebuilding it.
Where is our material processed?
The contract describes what Marram processes and where. Your material is not used to train AI models.