There is a specific moment in a company's life when institutional memory breaks, and it is more predictable than founders expect.

It is not the first year, when everyone who works there was in the room for every decision. It is not the fiftieth hire, though that is where people usually look for it. It is roughly years two through four — the stretch after the Series A, when the earliest employees start reaching the end of their equity vesting, the headcount has turned over enough that the majority joined after the decisions that shaped the product, and nobody has yet had a reason to write any of it down.

By then a startup is running on explanations that only four people can give, and two of them are leaving.

The Shape of the Attrition

An analysis published by Pin in 2026, drawing on more than four million observed job changes, found that 29.6 percent of startup hires leave within their first twelve months and 58.9 percent are gone within twenty-four. Median tenure at venture-backed startups ran 13 to 16 months depending on the funding stage at which someone was hired, peaking at Series A.

That is vendor-published analysis built on inferred career-transition data rather than payroll records, so it deserves a caveat — but the shape it describes is consistent with what harder sources show around it.

The Bureau of Labor Statistics found median employee tenure across all U.S. wage and salary workers at 3.9 years as of January 2024, the lowest reading since 2002. For workers aged 25 to 34 — the band that dominates startup headcount — it was 2.7 years. In professional and business services it was 3.5.

An older but mechanically instructive study from Carta, published in 2018 on 185,000 equity-holding employees across more than 3,600 companies, identified the driver: voluntary departures show a discontinuous jump right at the one-year vesting cliff, and by year four voluntary termination reaches 44.4 percent against 12.3 percent involuntary. That data is eight years old and predates two complete hiring cycles, so it should not be read as current — but the vesting-cliff mechanism it documents has not changed, because the vesting schedule hasn't.

Why Year Three Specifically

Run the arithmetic on a normal scaling trajectory.

A company is 15 people at seed. It raises a Series A and goes to 60 over the following two years. Some of the original 15 leave — normally, without drama, around their cliff or their four-year mark. By month thirty the company has 60 employees, maybe 8 of whom predate the Series A.

That means something specific: 87 percent of the company was not present for the decisions that determine how the company works. Why the pricing model is structured this way. Why the enterprise tier exists and what the deal was that forced it. Why the product deliberately doesn't do the obvious thing a competitor does. Which market segment was tried and abandoned, and what it cost to learn that.

Nobody wrote it down, because at the time everybody knew. That is the trap. Documentation feels redundant in a 15-person company — and it is redundant, right up until it isn't, at which point the window has closed and the people who could have written it are gone or too busy.

The Current Environment Makes It Sharper

The 2024–2026 conditions have intensified this in a way that is worth being concrete about.

Carta's data on its own platform shows startup hiring has normalized sharply from the boom. In January 2026 startups made 26,030 new hires against 20,378 departures — a hire-to-departure ratio of about 1.3, down from 3.8 at the January 2022 peak. Voluntary departures were down 40 percent from the mid-2022 high and layoffs down 57 percent from the January 2023 peak.

SignalFire's 2025 State of Talent report, citing Carta data, found Series A tech startups running about 20 percent smaller than their 2020 equivalents. New graduates now make up under 6 percent of startup hires, down more than 30 percent from 2019 levels.

The picture that adds up to: teams are leaner, hiring is more deliberate, and companies are leaning on experienced hires rather than training juniors. Every one of those shifts raises the cost of a slow ramp. A 60-person company with no slack cannot absorb a new senior hire spending six weeks reconstructing why things are the way they are — and an experienced hire is exactly the person who will change something before understanding it, because changing things is what they were hired to do.

Add the funding environment. PitchBook data reported in August 2025 put down rounds at 15.9 percent of venture deals, a decade high. Longer gaps between rounds mean longer periods where a company must execute on the strength of what it already knows, with fewer people, and no tolerance for relearning what it has already learned once.

What Gets Lost, Specifically

The losses at this stage are not customer records or code. Those survive. What evaporates is the layer of reasoning underneath them:

Why the product doesn't do the obvious thing. Every maturing product has three or four conspicuous gaps that look like oversights and are actually decisions. A new engineer or PM who doesn't know the difference will either rebuild something that was deliberately rejected or refuse to touch something that is genuinely just old.

What the first enterprise deal cost you. The commitment made to close it, the architectural concession, the customer who now has terms nobody else has and a contract nobody has reread.

The segment that didn't work. Startups reliably re-enter a market they already tried, because the attempt lives in a Slack retro from 2024 and the three people who remember it have left.

Pricing archaeology. The current price is almost never the result of a single decision. It is the residue of four, and only one of them was written down.

Who the load-bearing customers are and why. Which logo is in the deck because it's strategic, which one is genuinely large, and which one is quietly about to churn.

Why a particular technical constraint exists. The workaround built to bridge two systems, still running, understood by nobody currently employed.

The Usual Answer Doesn't Work at This Stage

The standard prescription is to write more things down. Institute an RFC process. Keep a decision log. Run better retros.

That advice is correct and it arrives too late. A company at 60 people can start documenting decisions from now on. It cannot retroactively document the previous three years, because doing so would require the specific people who are leaving to sit down and produce, from memory, an account of everything they decided — during their notice period, which they will spend transitioning their actual responsibilities.

This is the same structural failure that shows up in every context: the mechanism routes through a busy person's memory at exactly the moment that person is least available.

But the premise underneath the standard advice is wrong in a useful way. It assumes the reasoning wasn't recorded. In fact almost all of it was — in the Slack thread where the pricing was argued, in the pull request comment explaining the workaround, in the doc that proposed the segment and the one that quietly stopped being updated, in the CRM note from the enterprise deal, in the ticket where the constraint was first hit.

Startups are, if anything, over-recorded. They are simply unassembled. The record exists in fragments across six tools, none of which was designed to answer the question a new hire actually has.

Synthesis Is the Retroactive Version

What a scaling company needs is not better documentation discipline going forward — though it should have that too. It needs someone to read the last three years and write the account that was never written.

That is a reading problem, and reading is the thing software can now do at a scale a person cannot. Synthesis means connecting read-only to the systems that hold the record — the CRM, the Slack history, the docs, the issue tracker — and producing a document per new hire covering the decisions that shaped their area, the reasoning behind them, what was tried and abandoned, and who was involved, with a source link on every claim so the reader can open the original thread, see the date, see who was in it, and notice that the decision was made by someone who has since left.

That last detail matters more at a startup than anywhere else, because at a startup the answer to "is this still true?" is frequently no. A briefing that asserts things without provenance is worse than nothing when the underlying reality has moved twice since the thread was written.

The Window

The uncomfortable thing about year three is that it is the last cheap moment.

While the early employees are still there, the record can be assembled and checked against people who were present. Once they're gone, the record is all there is — still readable, still valuable, but with nobody left to correct it.

Most companies discover the problem on the other side of that window, usually when a new VP asks why something works the way it does and finds that the honest answer is that nobody currently employed knows.

The company hasn't lost the information. It has just never read it back.

See Firstweek pricing and plans.

Related: What Is Sales Context Synthesis? for the underlying concept, and How to Onboard a New RevOps or Marketing Ops Hire Without Losing Process Context — the same decision-provenance problem in a specific function.

Sources: Pin (Love Thy Recruiting, Inc.), "Employee Tenure by Funding Stage: 2026 Study of 4M+ Job Changes," July 2026 — 29.6% of startup hires leave within twelve months, 58.9% within twenty-four; median tenure of 13–16 months by stage at hire. Note: this is a vendor-published analysis of a proprietary profile index using inferred career transitions rather than payroll or HR-system records, and should be read as directional. U.S. Bureau of Labor Statistics, "Employee Tenure in 2024" — median tenure of 3.9 years across all wage and salary workers, the lowest since January 2002; 2.7 years for workers aged 25–34; 3.5 years in professional and business services. Carta, "Employment Tenure Statistics for Startup Companies," March 2018 — voluntary termination reaching 44.4% by year four against 12.3% involuntary, with a discontinuous increase in voluntary departures at the one-year vesting cliff, across 185,000+ equity-holding employees at 3,600+ companies. Note: this study is from 2018 and is cited here for the vesting-cliff mechanism, not as current data. Carta, "State of Startup Compensation: H2 2025" — 26,030 startup hires against 20,378 departures in January 2026, a ratio of roughly 1.3 against 3.8 at the January 2022 peak; voluntary departures down 40% from the June 2022 peak and layoffs down 57% from the January 2023 peak. Note: Carta's dataset covers companies using its cap-table platform and skews toward venture-backed, equity-issuing startups. SignalFire, "The SignalFire State of Talent Report – 2025" — Series A tech startups approximately 20% smaller than in 2020 (citing Carta data), and new graduates comprising under 6% of startup hires, down more than 30% from 2019. PitchBook data reported by Yahoo Finance, August 2025 — down rounds at 15.9% of venture-backed deals year-to-date, a decade high; cited via secondary reporting rather than PitchBook's own page.