90% of the rules you set with your AI are forgotten after 5 memory compressions.

That's the finding of a University of Passau study on AI agent memory, released a few weeks ago.

We all live this: when a conversation gets too long for the model's context, the AI compacts it. It summarizes its own history to make room. You would expect the model to preserve your most important rules each time, right? Well, no.

The study measured what survives:

- After 1 compaction: only 53% of the safety rules and constraints given to the agent are still intact.
- After 5 compactions: only 10% remain intact.

The instruction you gave on Monday ("never contact customers directly", "always price in EUR") has a 9-in-10 chance of being gone by Friday.

Why rules die first: they are short, stated once, and often phrased like facts. To the model, that looks like droppable background. Your history of small talk survives; your non-negotiables don't.

Here is how to protect your non-negotiables in your chats today:

1/ Never trust the transcript. Anything important said once in a long conversation is on a countdown.

2/ Pin your rules outside the compaction path. Put them in project instructions, claude.md and skill files — the layer your AI re-reads fresh every session — not in the flow of conversation.

3/ Write rules as rules. "Invoices are approved by the CFO" reads like a fact and gets compressed. "Never send an invoice without CFO approval" survives — atomic, explicit, one rule per line (even better if pinned, see point 2).

But the real problem goes beyond the chat window.

Right now, your company's AI memory is scattered across hundreds of individual chat sessions — each one compacting, each one forgetting, each one starting again from 10% next Monday. Your decisions, your constraints, your ways of working: rewritten as summaries of summaries, or simply lost. That is not a memory. That is collective amnesia.

Now, to become a competitive asset, AI memory has to be developed and managed at the company level. If this topic is on your agenda, let's talk!