Engineers know technical debt: the compounding cost of yesterday's shortcuts. Finance knows interest: the price of money you have already spent. Every company now carries a third debt, and almost nobody has it on a dashboard.
Knowledge debt is the recurring cost of relearning what your organisation already learned.
Every lesson your company holds was paid for once, and usually paid for dearly. The pricing structure that works was preceded by two that failed. The supplier you trust was found after one who burned you. The architecture decision that looks obvious in hindsight consumed a quarter of arguing. Salaries, failed experiments, lost deals, and lately, tokens: that was the purchase price of your company's judgement.
When the lesson is not kept, the purchase price is not the end. It is the first instalment.
The second instalment falls due when the person who learned it leaves. The third, when a new team faces the same question and runs the same experiment. In a human-scale company this waste was real but slow: institutional forgetting moved at the speed of staff turnover, reorgs and retirements.
Agents compound the debt at machine speed
Now add agents. Thousands of them, spun up fresh for every task, each with a context window that starts empty. Every one of them re-answers questions the company has already paid to answer. Every one re-derives context some other mind assembled this morning, token by token, at full price.
This is the part that should worry anyone signing an AI budget: retrieval-based architectures pay the reconstruction cost on every single query. Store the artefacts, retrieve some fragments, ask a model to rebuild the context. The archive grows, so the reconstruction gets more expensive precisely as the company gets older and wiser. Your inference bill is not just compute: a growing share of it is interest on knowledge debt.
Run the numbers on a 500-person company with a few thousand active agents and the figure lands somewhere that sounds made up: trillions of agent actions a year, a meaningful fraction of them re-purchasing knowledge the company already owns.
Paying it down
The fix is the same one accounting applied to every other invisible cost: put it on the books, then stop incurring it. An organisational memory that captures the trajectory of work as it happens, links it with provenance, and hands it to the next authorised mind means the lesson is bought once and inherited forever.
The economics invert. Instead of the bill rising as the archive grows, it falls as the company learns, because the brain loads the relevant thread instead of reprocessing the corpus. Memory becomes the one AI investment with compounding returns: every decision made today is a cost your company never pays again.