The memory you need to build for AI

By Janus Boye

“Every organisation runs on decisions nobody wrote down.”

— Andrei Kalamkarov

Andrei Kalamkarov leads the team behind Concordia University’s web platform

AI works well when it understands the context in which it is operating. Yet much of that context has never been written down.

In a recent member call, Andrei developed ideas he first shared in The memory I’m building for AI before CMS Connect 26.

He leads the team behind Concordia University’s web platform and has worked with its CMS since the beginning.

Mature systems carry years of decisions, exceptions and compromises. Without that history, AI may confidently replace something unusual with standard practice. The correction may be technically sound and still break something that was done deliberately.

Concordia’s web platform has been running since 2013. It now supports around 98,000 pages, more than 170 custom components and over 300 content authors. No single person can hold the history of a system that large in their head.

One small example captures the difficulty:
For years, authors could enter a single space in a title field to prevent a heading from appearing. When that whitespace was later removed, hundreds of headings reappeared. Removing the space looked like an obvious improvement. What the system could not explain was that the space carried meaning.

The same problem exists beyond code. Policies, content, workflows and organisational structures also carry decisions whose original reasoning may no longer be visible.

Memory is more than documentation

Andrei’s response was not simply to fix the problem. His team also recorded the hidden rule where AI would encounter it during future work.

This points to a form of organisational memory that goes beyond conventional documentation. It includes the unusual decisions that remain important, the exceptions that should be preserved and the approaches that have already been considered and rejected. Crucially, it records why those choices were made.

Andrei has developed this thinking into the Greenhouse, a shared memory system connecting project plans, decisions, reference material and reusable AI workflows. The technology is deliberately ordinary: Markdown, Obsidian, Git and GitHub, with selected material published to Confluence.

At CMS Connect 26, Matt Garrepy from CMS Critic gave this broader shift a useful name:

“Memory is becoming infrastructure.”

Andrei returned to that observation throughout the member call. The Greenhouse shows what it can mean in practice.

The important part is not the choice of tools. It is making the organisation’s accumulated judgement available where people and AI are doing the work.

Build memory into the work

The Greenhouse is designed as a working environment rather than a finished archive. Andrei’s team separates its contents into three layers: the Stream for working notes and meeting summaries, Raw for AI-generated material that has not yet been validated, and the Garden for refined knowledge that has been reviewed and is worth reusing.

That distinction matters because AI makes it possible to generate far more documentation than before. More information does not automatically create better organisational memory. Unreviewed output may create the appearance of knowledge without providing anything people should trust.

Andrei demonstrated how this memory can support everyday work. One workflow reads technical changes and project information, then drafts release notes for Concordia’s content authors. Work that previously required about an hour of drafting can now be completed with a few minutes of human review.

The time saving is useful, but the larger change is that planning, development and communication are working from the same memory.

More context is not always better

It is tempting to solve the problem by giving AI as much information as possible. Andrei’s experience suggests that this is not enough.

Long lists of instructions can conflict, become difficult to apply and distract from the immediate task. Context needs to be selected and organised rather than simply accumulated.

During the call, Mirco Fabris, Head of Technology Onboarding at ctrl QS, a Berlin-based agency and Boye & Company member, connected this to a demonstration at Google Builders 2026 in Berlin, where performance gradually deteriorated as more context was added. As he observed during the call, this is not only a coding problem. It applies wherever organisations expect AI to work from a growing body of knowledge.

Concordia handles this by keeping broad principles at the top level while loading more specific information for individual projects and tasks. A useful instruction is clear about when it applies and explains why the underlying decision was made.

“Sound human” offers little practical guidance. An instruction to check whether a field is intended to contain HTML before changing how it is displayed can be applied at a specific moment. Adding the reason behind a rule also helps AI navigate situations where different instructions pull in different directions.

It is equally valuable to record rejected paths. If an organisation has already considered an approach and decided against it, AI should not repeatedly present it as a new recommendation.

Memory requires ownership

A shared memory system does not remove the need for judgement. It makes that judgement more visible and reusable.

During the member call, Maarten Korz, Innovation leader at ZF Group asked whether the refinement stage in the Greenhouse always requires human involvement. For Andrei’s team, it does. AI can capture, organise and draft, but a person decides what becomes trusted organisational memory.

Soeren Stamer raised another important question: should teams document every legacy exception or remove the technical debt instead? In the case of the single-space title, Concordia ultimately did both. The team recorded the hidden behaviour so it would not be broken again, then cleaned up the affected content.

Memory should not become an excuse to preserve every historical choice forever. Some decisions remain necessary. Others have outlived the conditions that created them.

The same applies to the memory itself. AI makes adding information easy, while removing outdated material still requires deliberate effort. An obsolete instruction can be more dangerous than a missing one because AI may treat it as current fact.

Someone therefore needs to tend the memory: reviewing what enters, questioning what remains and removing what no longer reflects reality.

Institutional memory is becoming infrastructure

Every mature organisation has its equivalent of Concordia’s single space: a choice that looks accidental but carries hidden history.

Until now, organisations have often relied on particular people to remember why such choices were made. AI makes the weakness of that arrangement more visible. It cannot preserve reasoning it has never been given.

Building memory for AI does not mean documenting everything. It means identifying the decisions that continue to shape the work and making their reasoning available when it matters.

The challenge is not only what the organisation needs to remember. It is who will maintain that memory, who will question it and who will know when it is time to let something go.

The conversation continues

If reading and being a part of an online call is not enough, you’re very welcome to join us more actively. Our community is built around learning together, comparing notes, and exploring how theory meets practice across roles, industries and regions.

You can also:

However you choose to engage, we’re glad you’re here and part of the journey.