The problem

The agentic enterprise is here.
Its memory is not.

Your organisation already runs on thousands of minds, human and artificial. What it lacks is any way for those minds to inherit what the others have learned.

500 humans. 2,000 personal agents. 5,000 task agents. Trillions of agent actions a year.

Machine traffic already exceeds human traffic. Cloudflare forecasts 1,000x within five years.

Fifty years of systems of record

Enterprise software was built for static information.

Every generation of enterprise software has answered the same question: where do we put the artefacts? Documents, databases, files, search. The ERP stored transactions, the CRM stored contacts, the wiki stored pages. All of it assumes the valuable thing is the output.

Today's enterprise AI keeps that architecture and bolts retrieval on top: store the artefacts, retrieve some fragments, ask a model to rebuild the context. Every single time. The model reconstructs, guesses at intent, and throws its reconstruction away the moment the answer ships.

MODELrebuilding context
DOCUMENTS · DATABASES · FILES · SEARCH
QUEUE
0
Then the swarm arrives

It breaks when thousands of agents act at once.

Retrieval-on-top works while software sits still and a human waits politely for an answer. Agents don't wait. They query in parallel, thousands of times a minute, and every query pays the full price of rebuilding context that some other mind already built this morning.

The result is an architecture doing something it was never designed for: serving as the working memory of an organisation, one expensive reconstruction at a time.

Knowledge debt

You already paid to learn this.

Every lesson your company has ever learned was paid for once: in salaries, in failed experiments, in tokens. When the lesson isn't kept, you pay for it again. That recurring, invisible cost is knowledge debt, and in an agentic company it accrues at machine speed.

0
tokens a year spent figuring out what your company already knew
Two kinds of IP

Your systems store the first. The second evaporates.

THE FIRST

The work you ship

Code. Products. Documents. Data. Safely stored, for fifty years.

THE SECOND

The judgement behind it

Why you decided. What you tried. What failed. What you learned. Gone by the next morning, every single day.

The second kind is the one competitors cannot copy, the one new hires need most, and the one no system of record has ever captured. It lives in heads, in threads, and in chat histories that expire.

The amnesia montage

Brilliant minds. Institutionally amnesiac.

Four frontier models, working the same problem in the same company, each burning the same tokens to reach the same conclusions. Watch the meters: the work is identical.

Claude
Claude
0
TOKENS
ChatGPT
ChatGPT
0
TOKENS
Codex
Codex
0
TOKENS
Copilot
Copilot
0
TOKENS
DUPLICATING WORK
CONTRADICTING DECISIONS
REBUILDING CONTEXT
ASKING AGAIN
FIGHTING THE COMPANY'S GOALS
FORGETTING WHY
The intelligence of the parts does not guarantee the intelligence of the whole.

A company full of brilliant agents can still behave stupidly as a whole. Each agent optimises its own task, and no one carries the thread.

Next
The platform: White Tree and Dúnedain
→