Contents
Computer Science & Systems Thinking
Section 1
Core Idea
Belady's algorithm — sometimes called MIN, or the clairvoyant algorithm — is the theoretical optimum for cache replacement. Published by László Bélády in 1966 while he was at IBM Research, the rule is stated in a single sentence: when the cache is full and a new item must be admitted, evict the item that will not be used for the longest time in the future. Given a finite cache and a known reference sequence, no algorithm can produce fewer misses than Belady's. It is optimal by construction, and it is proved optimal, which is a rare combination in systems literature.
It is also, in general, impossible to implement. The rule requires knowledge of the future reference sequence — which items will be requested next, and in what order. Real systems don't have that information. What Belady's algorithm gives us is not a runnable procedure but a theoretical upper bound: the number of cache misses any real replacement policy achieves can be compared against MIN's miss count on the same reference trace, and the gap is the tax that ignorance of the future is charging. Every real cache policy — LRU (least recently used), LFU (least frequently used), CLOCK, ARC, 2Q — is an approximation of MIN, each making a different guess about which past behaviour best predicts future references.
The intellectual weight of the model is not in the algorithm itself. It is in the reframing. Once you accept that the optimal replacement rule is "evict what won't be needed longest," every real system's suboptimality becomes measurable, and every heuristic becomes explainable as a specific bet about how the future resembles the past. LRU assumes recency predicts imminence. LFU assumes frequency predicts persistence. They can both be right in different workloads, and they are both wrong in the presence of a rare but exact predictor. Belady's contribution was not to make caches faster; it was to make the theoretical structure of the problem legible.
For an operator, the model's business analogues are almost embarrassingly direct. Attention is a cache. Calendars are caches. Team focus is a cache. Portfolios are caches. In each case, the number of slots is finite, the incoming demand is unbounded, and the eviction rule you use — whether explicit or accidental — determines the miss rate. Belady tells you what the ceiling looks like even when the ceiling is unreachable.
Section 2
How to See It
Attention & Focus
You're seeing a Belady problem when you're deciding which of many active priorities to drop, and the natural heuristic ("what have I not touched recently") is producing evictions that later cause expensive reintroduction cost. LRU on your calendar is often worse than a Belady approximation that considers future touch frequency.
Portfolio Management
You're seeing it when a portfolio needs to be trimmed and the tempting rule ("cut the losers, keep the winners") is a variant of LRU applied to price momentum rather than to expected future usefulness. Belady's frame asks a different question: which position will contribute least to the future return path?
Product & Roadmap
You're seeing it when a product team has to drop features to hit a release date. Cutting the features nobody has looked at recently is LRU-style eviction; the correct question is which features will be least needed in the next customer cohort — a Belady-shaped estimate that requires forecasting, not backward-looking data.
Section 3
How to Use It
Belady's algorithm is useful in a business setting not as a procedure to run but as a discipline of question. The wrong question is "what have I done least recently?" The right question is "what will I need least soon?" That reframe is uncomfortable because the second question requires forecasting and the first does not, and forecasting is harder. But every high-leverage attention decision an operator makes is a Belady-shaped decision in disguise, and using LRU by default is leaving performance on the table.
Decision filter
"If I could see one week into the future, which of my current commitments would I be most relieved to have dropped? That is the item Belady's algorithm would evict."
The practical procedure has three parts. First, define the cache. How many slots are you actually operating with — meeting slots per week, active projects, portfolio positions, direct reports? Number the finite resource honestly; overload it and every rule collapses into thrashing. Second, forecast the reference sequence. For each item in the cache, estimate when it will next be genuinely needed — not "when did I last touch it," but "when will future me be relieved I kept it." Third, evict the one with the most distant expected reference. In practice you'll be uncertain, so the honest approximation is to evict items with high uncertainty and low expected forward utility, rather than items with high recency alone.
As a founder or operator
Apply Belady quarterly to your active project portfolio. List every workstream currently consuming meaningful attention. For each, estimate the next date it will produce a decision that materially changes the business. Kill or defer the workstreams whose next-decision date is furthest into the future — not the ones you've touched least. This is why "cut the small projects" is often a mistake; a small project with a near-term decision is more valuable in the cache than a large project whose next decision is a year out.
As an investor
When trimming a portfolio, use Belady's logic against your own trailing bias. The typical instinct — cut losers, keep winners — is LRU on price. The Belady question is: which position is least likely to matter in the next return path? A recent loser that is one catalyst away from doubling is a high-utility cache item; a flat position with no near-term catalyst is a lower-utility item even if it has been quietly compounding. The exercise almost always produces a different sell list than the price-momentum heuristic.
Section 4
Common Misapplications
Three failure modes recur.
First, overloading the cache. The number one cause of Belady-suboptimal outcomes in real business systems is not the wrong eviction rule; it is admitting more items than the cache can hold in the first place. A founder running twelve active projects, an investor holding forty concentrated positions, a manager overseeing fifteen direct reports — all are past the point where any replacement algorithm can help. Cut the cache size first, then argue about the eviction rule.
Second, confusing LRU with Belady. LRU is the most common approximation, and in many workloads it is excellent, because recency is genuinely correlated with imminence. But LRU fails catastrophically in workloads with cyclic references — where an item is used, unused for a long stretch, and then needed again on a predictable schedule. Annual planning cycles, quarterly board prep, seasonal customer patterns — these are cyclic reference workloads, and LRU-style eviction produces the exact wrong answer. Belady's rule would keep the cyclic item; LRU would evict it just before it's needed.
Third, treating the theoretical bound as a practical target. Belady's algorithm is optimal only under perfect foresight. Real operators cannot achieve it, and pretending they can — by over-forecasting confidence in which items will matter — is worse than admitting the uncertainty and running a well-tuned approximation. Know the bound exists, measure how far your current heuristic is from it, and invest in better forecasting only where the gap is expensive enough to justify the effort.
There is also a cognitive misuse worth naming. Belady rewards forecasting, and operators who instinctively resist forecasting will resist the frame. That resistance is often a tell that the cache is being run inefficiently — the reluctance to name future dates for present items is the same reluctance that makes LRU the default.
Section 5
Founders & Leaders
Charlie MungerVice Chairman, Berkshire Hathaway
Munger's disciplined refusal to hold more than a handful of concentrated positions is a Belady-optimal cache design. He operates with a very small cache and evicts aggressively based on forward expected utility, not backward performance. The "sit on your ass" investing style is the discipline of not admitting new items into the cache unless they meaningfully improve the forward reference sequence — a strict form of Belady applied to a lifetime of capital allocation.
Steve JobsFounder, Apple
Jobs's product-portfolio pruning on his return to Apple in 1997 is the cleanest historical case of a Belady-style eviction on a corporate scale. Apple had dozens of overlapping product lines. Jobs collapsed them to a four-quadrant grid — pro vs consumer, desktop vs laptop — evicting everything else based on which products were unlikely to be needed in the near-term reference sequence of a focused company. The pruning was uncomfortable, unpopular, and correct.
Section 6
Company Examples
Netflix
Netflix's content library is a real-world Belady cache. Every title costs money to license and takes attention to surface; the total slot count is finite. Netflix's content decisions increasingly resemble a Belady approximation on future viewership rather than LRU on past viewership — cancellations of shows with declining forward-audience curves happen faster than legacy usage would justify, precisely because the operating question is which title will contribute least to future subscriber retention.
Amazon
Amazon's fulfilment-centre inventory placement is a large-scale Belady approximation. The system does not stock what customers recently bought at each warehouse; it stocks what it forecasts they will soon need there. The gap between LRU-style placement (recent sales) and Belady-style placement (forecast demand) shows up directly in delivery times and margins, which is why the forecasting infrastructure is one of the company's least-visible but highest-leverage assets.
Section 7
Connected Models
Applies
Opportunity Cost
Every cache slot is an opportunity cost. Belady's algorithm is opportunity cost applied to a finite, replaceable resource with a known constraint — the sharpest possible form of the general model.
Counteracts
Sunk Cost Fallacy
The instinct to keep a workstream because you've invested in it is LRU corrupted by history. Belady's rule ignores what you've spent and asks only about future utility — the correct frame for eviction.
Requires
Second-Order Thinking
Optimal eviction depends on forecasting the downstream cost of reintroduction. An item cheap to bring back next month is different from an item that would take a quarter to rebuild, and Belady applied naively ignores that reintroduction cost.
Pairs-with
Optimal Stopping
Both models formalise finite-resource allocation under uncertainty. Belady tells you what to evict; optimal stopping tells you when to admit a new candidate. Together they describe the whole cache lifecycle.
Section 8
One Key Quote
— László Bélády, A Study of Replacement Algorithms, IBM Systems Journal, 1966"The optimum algorithm replaces that page whose next reference is farthest in the future. Since this requires foreknowledge of the reference string, the algorithm cannot be realised; it serves rather as a yardstick against which realisable algorithms may be measured."
Section 11
Summary & Further Reading
Belady's algorithm is optimal, impossible to implement, and quietly one of the most useful mental models for anyone managing a finite resource — attention, capital, calendar, headcount, product surface area. It reframes the problem of eviction away from "what have I touched least?" toward "what will I need least soon?" and it makes the difference between the two measurable. Every real system operates with an approximation of MIN; the operators who compound are the ones who know they are approximating and invest in the forecasting that closes the gap.
01
Paper
The original IBM Systems Journal paper. The proof of optimality and the definition of the MIN algorithm remain the foundation for every modern replacement policy — a piece of applied theory that has aged unusually well.
02
Book
The best popular treatment of caching, optimal stopping, and other computer-science algorithms as models for everyday decision-making. Chapter 4 covers Belady, LRU, and their behavioural analogues directly.
03
Book
The applied manual for running a small, disciplined attention cache. McKeown does not use Belady's language, but the operating philosophy — evict aggressively based on forward priority, not on legacy commitment — is the same model in plain English.
Why this matters next
mental modelsSecond-Order Thinking
Belady's Algorithm applied the Second-Order Thinking mental model
mental modelsLeverage
Belady's Algorithm applied the Leverage mental model
mental modelsCompounding
Belady's Algorithm applied the Compounding mental model
mental modelsMomentum
Belady's Algorithm applied the Momentum mental model
mental modelsSystems Thinking
Belady's Algorithm applied the Systems Thinking mental model
mental modelsBelady's Algorithm
Belady's Algorithm applied the Belady's Algorithm mental model
Frequently asked questions
What is Belady's Algorithm?
Belady's Algorithm is a mental model used for better thinking and decision-making.
How do you apply Belady's Algorithm?
To apply Belady's Algorithm, identify situations where this framework is relevant, then use it as a lens to evaluate your options and decisions. The model is most useful when combined with other complementary mental models.
What category does Belady's Algorithm fall under?
Belady's Algorithm falls under the Computer Science & Algorithms category of mental models. Other models in this category can be found on the Computer Science & Algorithms hub page.
Why is Belady's Algorithm important?
Belady's Algorithm is important because it provides a structured way to think about problems that would otherwise be approached with intuition alone. Understanding this model helps you avoid common reasoning errors and make better decisions.
Where does Belady's Algorithm come from?
Belady's Algorithm is discussed in the tradition of Belady.
Continue exploring
ABETMLMLMMTD
Mental model
Abstraction
The practice of hiding complexity behind a simpler interface, enabling reasoning
Mental model
Explore-exploit Tradeoff
The fundamental tension between gathering new information (exploring) and levera
Mental model
Metcalfe's Law
The value of a network grows proportionally to the square of the number of its u
Mental model
Moore's Law
The observation that transistor counts on integrated circuits double roughly eve
Mental model
Mythical Man Month
Brooks's Law: adding people to a late software project makes it later due to com
Mental model
Technical Debt
The accumulated cost of expedient decisions in software and systems that must ev
More like this, in your inbox
I send a newsletter every week — free, no spam, unsubscribe anytime.