Nobody remembers why
In 2011, software architect Michael Nygard published a short essay titled "Documenting Architecture Decisions" that diagnosed a problem plaguing every software team he had ever worked with. Teams would inherit codebases full of decisions they did not understand. Why was the system built this way? Why was this technology chosen over that one? Why does this constraint exist? The answers were gone — dissolved into the memories of people who had left the company, buried in Slack channels nobody would ever scroll back to, or simply never written down at all (Nygard, 2011).
Nygard's solution was a lightweight document format he called an Architecture Decision Record, or ADR. Each ADR captured five things: a title, a status, the context surrounding the decision, the decision itself, and the consequences. The critical innovation was not the decision or the consequences — teams already tracked those, at least informally. The innovation was the context section. Nygard insisted that every ADR explicitly record the forces, constraints, requirements, and trade-offs that were active at the moment the decision was made. Not what the team decided. Why the team decided. What pressures were they under? What did they know? What did they not know? What alternatives did they reject, and for what reasons?
The format was deliberately short — one to two pages. It was stored alongside the code it described. And it was written, as Nygard put it, as if in conversation with a future developer. The whole point was to make the motivation behind previous decisions visible to everyone, present and future, so that nobody would be left scratching their heads asking "What were they thinking?"
ADRs spread across the software industry because they solved a problem that every team recognized: the slow, silent disappearance of decision context. But the problem Nygard identified is not unique to software. It is universal. Every decision you make — personal, professional, strategic, trivial — has a context that will evaporate unless you capture it in writing. And when that context disappears, misinterpretation is not a risk. It is a certainty.
Your memory rewrites the past
The reason written context is necessary is not that memory is imperfect. It is that memory is actively deceptive.
In 1975, psychologist Baruch Fischhoff published the foundational research on what he called hindsight bias — the systematic tendency of people to believe, after learning the outcome of an event, that they "knew it all along." Fischhoff presented subjects with historical scenarios and asked them to estimate the probabilities of different outcomes. Subjects who were told the actual outcome consistently rated that outcome as more likely than subjects who were not told — and they were largely unaware that knowing the outcome had changed their estimates (Fischhoff, 1975).
The implication for decision context is devastating. Once you know how a decision turned out, your memory of why you made it shifts to accommodate the outcome. If the decision worked, you remember being confident and well-reasoned. If it failed, you remember having doubts that you ignored. In both cases, the memory is a reconstruction — a story your brain assembles after the fact to create a coherent narrative — not an accurate record of what you actually thought and felt at the moment of choice.
Elizabeth Loftus's research on memory reconstruction deepened this finding. Loftus demonstrated across decades of experiments that memory is not a recording device — it is a reconstruction engine. In her landmark studies, she showed that post-event information — a leading question, a suggestive detail, even the verb used to describe an event — physically alters the stored memory itself. Subjects who were asked how fast the cars were going when they "smashed" into each other reported higher speeds and even "remembered" broken glass that was never present in the original footage (Loftus & Palmer, 1974). The original perception and the subsequent suggestion merged into a single memory that subjects could not decompose back into its parts.
This is what happens to the context of every unrecorded decision. The original forces, constraints, uncertainties, and emotions that shaped your choice do not sit patiently in memory waiting to be retrieved. They are overwritten by subsequent events. They are distorted by outcomes. They merge with later rationalizations until the original context is irrecoverable. You do not forget the context of your decisions. Your brain replaces it with a plausible fiction and presents that fiction as memory.
This is why you cannot reconstruct decision context from memory, no matter how vivid the memory feels. The only reliable preservation mechanism is writing — capturing the context at the moment of decision, before hindsight has a chance to rewrite it.
The organizational cost of lost context
The consequences of lost decision context scale with the number of people affected. When you forget why you made a personal decision, you waste your own time. When an organization forgets why a policy, architecture, or strategy was chosen, the waste compounds across every person who interacts with that decision.
Research on institutional memory loss shows a consistent pattern. When organizations experience staff turnover, restructuring, or system migrations, the formal records of what was decided typically survive — but the informal knowledge of why it was decided does not. The result is what researchers call institutional amnesia: a state in which the organization has lost the ability to understand its own reasoning (Pollitt, 2000). Decisions that were made for specific, defensible reasons are questioned, reversed, or repeated because the reasoning has evaporated along with the people who held it.
The UK government's Centre for Science and Policy identified four primary drivers of institutional amnesia: high staff turnover, constantly changing IT and data management systems, regular organizational restructuring, and a culture that rewards management skills over the ability to hold and recall knowledge. Notice that none of these drivers are about the quality of the original decision. They are all about the failure to preserve context across time and personnel changes. The decisions may have been excellent. Without their context, they become opaque artifacts that the organization must either trust blindly or investigate from scratch.
This dynamic explains a pattern that anyone who has worked in a large organization recognizes: the same debates recurring every two to three years as new leadership arrives, discovers existing policies, questions why they exist, and either spends months rediscovering the original reasoning or reverses the policy and re-encounters the original problem. The cycle is not caused by bad decision-makers. It is caused by the absence of written context connecting decisions to their rationale.
Decision journals: the personal ADR
If Architecture Decision Records solve context loss for teams, decision journals solve it for individuals.
Shane Parrish, founder of Farnam Street, popularized a structured decision journal practice in 2014 that applies the same principle to personal decisions. The format captures the state of your mind at the moment of decision: what you decided, what alternatives you considered, what information you had, what uncertainties you faced, and critically, what your emotional state was when you committed to a course of action (Parrish, 2014).
The emotional state component matters more than most people expect. When you review a decision six months later, you will remember the facts you considered. You will not accurately remember the fear, excitement, fatigue, or social pressure that shaped how you weighted those facts. The decision that looks irrational in retrospect may have been perfectly rational given the emotional forces that were active at the time — forces that are invisible in hindsight unless you wrote them down.
Annie Duke, in her work on decision quality, draws a sharp distinction between decision quality and outcome quality that depends entirely on preserved context. A good decision made with good reasoning can produce a bad outcome due to factors outside your control. A bad decision made recklessly can produce a good outcome due to luck. If you evaluate decisions only by outcomes — which is all you can do when the original context is gone — you will systematically reward luck and punish good reasoning. You will learn exactly the wrong lessons from your own experience (Duke, 2018).
The decision journal prevents this. When you record the context of a decision at the time you make it, you create a ground truth document that future evaluation can reference. Did the decision fail because the reasoning was flawed? Or did it fail because circumstances changed in ways you could not have predicted? These are different problems with different solutions, and you cannot distinguish between them without the original context.
Here is the format, synthesized from Parrish's journal structure and Nygard's ADR format, adapted for personal use:
Decision Context Record
- Date and decision: What you decided, in one sentence.
- Context: The forces active at the moment of choice — what you knew, what you did not know, what constraints you faced, what emotions you felt, what alternatives you considered and rejected.
- Expected consequences: What you believe will happen as a result, including the timeline.
- Confidence: A number from 1 to 10 representing your certainty.
- Review trigger: When you will revisit this record to compare expectations against reality.
This takes three to five minutes. The return on those three to five minutes is the ability to learn accurately from your own decisions for the rest of your life.
Writing externalizes cognition that memory cannot hold
The reason writing works where memory fails is not simply that writing is more permanent. It is that writing forces a different cognitive process — one that creates a more complete and coherent representation of context than memory ever could.
James Pennebaker's research on expressive writing, spanning four decades and hundreds of studies, demonstrates that the act of writing about experiences activates cognitive processes that do not occur during mere thinking or verbal discussion. When people write about events, they engage in causal reasoning — constructing explanations for why things happened. They use insight language — words like "realize," "understand," and "because" — at increasing rates over multiple writing sessions. The writing process itself generates coherence that was not present before the pen hit the page (Pennebaker, 2018).
This finding has a direct application to decision context. When you sit down to write why you made a decision, you are not simply transcribing a pre-existing mental record. You are constructing a more complete record than existed in your head. The act of writing forces you to make implicit reasoning explicit. It surfaces assumptions you did not know you were making. It reveals gaps in your thinking that were invisible when the reasoning was purely internal. The written context record is not a copy of your mental state — it is an improvement on it.
This is externalized cognition in its most practical form. Your working memory holds approximately four chunks of information simultaneously. The context of a significant decision involves dozens of factors — constraints, emotions, alternatives, uncertainties, time pressures, social dynamics, domain knowledge. No human working memory can hold all of these simultaneously. But a written document can. Writing does not just preserve context — it creates a representation of context that is more complete than what your brain could hold at any single moment.
AI and the Third Brain: context as the operating system
The parallel between human context loss and AI context loss is not a metaphor. It is a structural identity.
Large language models depend entirely on explicit context to function. An LLM has no persistent memory between conversations. Every interaction begins with a blank slate — the model knows only what is provided in its context window. If you do not provide the context of a decision, the model cannot reason about that decision. If you provide partial context, the model reasons about a partial picture and produces confidently wrong conclusions. The context window is not supplementary to the model's intelligence. It is the operating system on which that intelligence runs.
Research on what Chroma's 2024 technical report calls "context rot" demonstrates that even within a single conversation, LLM performance degrades as context accumulates. Models claiming 200,000-token context windows often become unreliable around 130,000 tokens, with sudden performance drops rather than gradual degradation. More critically, semantic similarity between the question and surrounding context material causes the model to confuse relevant context with irrelevant context, producing answers that sound authoritative but draw from the wrong information (Chroma Research, 2024).
This means that working with AI effectively requires the same discipline that this lesson teaches for human cognition: explicit, structured, well-organized context records. When you ask an AI to help you evaluate a past decision, the quality of its analysis depends entirely on the quality of the context you provide. If you provide a Decision Context Record — the forces, constraints, alternatives, and reasoning at the time of choice — the AI can genuinely help you analyze whether the reasoning was sound, whether the situation has changed, and whether the decision should be revisited. If you provide only the decision and the outcome, the AI will do exactly what your hindsight-biased brain does: construct a plausible narrative that connects the decision to the outcome and call it analysis.
The practical implication is that your written context records serve double duty. They protect your own reasoning from hindsight distortion. And they serve as high-quality input for AI-assisted decision review — turning your Third Brain into an analytical partner that can reason about your decisions with the same context you had when you made them.
The workflow looks like this: Write the Decision Context Record at the time of the decision. At the review date, provide the record to an AI system along with the actual outcome. Ask it to analyze the gap between expected and actual consequences. Ask it to identify which contextual factors you weighted correctly and which you misjudged. Ask it to flag any reasoning patterns that appear across multiple decision records. The AI does not have hindsight bias. It processes the original context and the outcome as separate inputs, which is exactly the separation your brain cannot maintain.
The context preservation protocol
Here is the deliberate practice that converts this knowledge into a persistent habit.
Step 1: Identify the decision threshold. Not every decision warrants a context record. The threshold is any decision where you consider more than one option for more than sixty seconds. If you deliberated, the context is worth preserving. If the choice was automatic, it was not really a decision — it was a habit or a default.
Step 2: Write the context at the moment of decision. Not after. Not tomorrow. At the moment you commit, take three to five minutes to record the five elements: decision, context, expected consequences, confidence, and review trigger. The context is freshest — and most honest — in the minutes immediately following the commitment.
Step 3: Store the record where it will be found. A context record that exists in a forgotten notebook or a buried file is functionally identical to no record at all. Store decision records in a dedicated, searchable location — a decision log file, a tagged note in your knowledge system, or alongside the artifact the decision produced (as Nygard recommended for ADRs stored next to the code they describe).
Step 4: Review on schedule. When the review trigger date arrives, read the original context record before evaluating the outcome. This sequence matters. If you evaluate the outcome first and then read the context, hindsight bias has already contaminated your assessment. Read the context first. Remember what you knew and did not know. Then look at what happened.
Step 5: Extract the learning. The gap between your expected consequences and actual consequences, read in light of the original context, is the raw material for genuine learning. Did the reasoning fail? Did the context change? Did you weight the wrong factors? These questions can only be answered honestly when the original context is preserved in writing.
The bridge to invisible context
Written context records solve the problem of temporal context loss — the erosion of decision reasoning over time. But there is a category of context that cannot be preserved by writing because it was never visible in the first place.
Cultural context — the assumptions, values, and interpretive frameworks that your culture installs in your perception without your awareness — shapes every decision you make. But unlike the constraints and trade-offs you can articulate in a Decision Context Record, cultural context operates below the threshold of conscious access. You do not know it is there until you encounter a culture that operates on different assumptions.
That is the subject of L-0165. The context loading you practiced in L-0163 taught you to deliberately reconstruct context when switching between domains. This lesson taught you to preserve context in writing so that reconstruction is possible months and years later. The next lesson confronts the hardest case: context that you cannot write down because you do not know it exists — until a cultural crossing makes it suddenly, sometimes painfully, visible.
Sources:
- Nygard, M. (2011). "Documenting Architecture Decisions." Cognitect Blog. Available at: https://www.cognitect.com/blog/2011/11/15/documenting-architecture-decisions
- Fischhoff, B. (1975). "Hindsight Is Not Equal to Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertainty." Journal of Experimental Psychology: Human Perception and Performance, 1(3), 288-299.
- Loftus, E. F., & Palmer, J. C. (1974). "Reconstruction of Automobile Destruction: An Example of the Interaction Between Language and Memory." Journal of Verbal Learning and Verbal Behavior, 13(5), 585-589.
- Pollitt, C. (2000). "Institutional Amnesia: A Paradox of the 'Information Age'?" Prometheus, 18(1), 5-16.
- Parrish, S. (2014). "How a Decision Journal Changed the Way I Make Decisions." Farnam Street. Available at: https://fs.blog/decision-journal/
- Duke, A. (2018). Thinking in Bets: Making Smarter Decisions When You Don't Have All the Facts. New York: Portfolio/Penguin.
- Pennebaker, J. W. (2018). "Expressive Writing in Psychological Science." Perspectives on Psychological Science, 13(2), 226-229.
- Chroma Research. (2024). "Context Rot: How Increasing Input Tokens Impacts LLM Performance." Available at: https://research.trychroma.com/context-rot