What does the forgetting curve mean for your next knowledge-work hire?

Forgetting research gives you a reason to support human recall, but it does not calculate which jobs AI should replace. Assess retention, relationships and tacit judgment before redesigning knowledge work.

On this page
  1. What did Ebbinghaus measure?
  2. How useful is the 40/35/25 framework?
  3. Why pay attention to mid-size companies?
  4. What this changes about the next hire
  5. Limits and counter-arguments worth naming

The short version

My argument is to assign recall-heavy tasks to AI and design human roles around relationships and judgment. That recommendation depends on checking the system against the actual work. The forgetting curve doesn't calculate which jobs to replace.

  • Ebbinghaus studied relearning, and Murre and Dros reproduced much of the pattern in a 2015 single-subject experiment. Those results are not onboarding-loss percentages.
  • The 40/35/25 split is a vendor framework for documented, informal and tacit knowledge. Use it as an example, then assess your own work.
  • Mid-size firms with little backup for key roles deserve attention, but this research does not rank their losses against other firms.
  • Ask how much of the role depends on recall, relationships and hard-to-explain judgment. Test AI on the actual tasks before changing the role.

You hire someone. You teach them the process once. That does not tell you what they can recall a week later. But the familiar claim that they lose 58% in twenty minutes, 75% in a day and 90% in a week does not follow from the Ebbinghaus evidence.

Don’t blame the hires. Hermann Ebbinghaus studied forgetting in experiments reported in 1880 and 1885. In 2015, Jaap Murre and Joeri Dros at the University of Amsterdam reproduced much of the pattern in PLOS ONE. My conclusion for employers is that memory deserves support, instead of an expectation that one round of training will last indefinitely.

My conclusion: where remembering rules and exceptions is central to a role, design it around AI doing that recall work and people handling the parts that need their judgment. A stored record can persist without biological forgetting. An AI still has to retrieve and use it correctly.

That doesn’t mean fire everyone. Empathy, reading a difficult conversation and deciding how to respond still need an accountable person. But remembering edge cases, recent decisions, last quarter’s exceptions, the contractor who burned you in 2022? Those are good candidates for records that people can search and check, with AI helping where it proves useful.

What did Ebbinghaus measure?

Ebbinghaus learned lists of nonsense syllables and measured the effort needed to relearn them later. He used himself as the subject over seven months. The material reduced the influence of familiar meanings, but it did not isolate a universal rate for all human knowledge. His savings score measured the proportion of original learning time saved on relearning. Zero savings means relearning took as long as the first attempt. It does not mean no trace of the material remained.

Murre and Dros’s 2015 replication, published in PLOS ONE, used one subject, Dros, for about 70 hours over a 75-day experimental phase. In Table 3 of the paper, his unadjusted savings scores were 0.472 after 20 minutes, 0.317 after a day, 0.168 after six days and 0.041 after 31 days. The pattern resembled Ebbinghaus’s, although Dros’s 31-day value was lower. After correcting for the increase in initial learning time during the experiment, the paper gives a 31-day savings estimate of 0.137.

That is the academic version.

The 58%, 75% and 90% figures used in corporate learning decks are not harmless rounding of those results. Savings in relearning time and the percentage of information someone can recall are different measurements. Neither the original experiment nor this replication measured how much a new employee forgot from onboarding.

The distinction matters when a memory experiment becomes a hiring argument.

Murre and Antonio Chessa published a follow-up online in 2022, in a 2023 journal issue, explaining why savings is a “pure” measure in a particular mathematical sense: it is independent of initial encoding strength and learning and relearning times. That does not prove a fixed biological rate of decay for every kind of memory.

The practical question is how much better we can make work by combining records, retrieval and AI assistance.

How useful is the 40/35/25 framework?

A 24G vendor article proposes a useful framework for tribal knowledge: 40% documented, such as SOPs and manuals; 35% informal, such as exceptions and “we don’t ship to that vendor on Fridays”; and 25% tacit, such as experience that is hard to explain. It supplies no research method for those percentages. Treat them as an example, not a measured split for a typical company.

Illustrative knowledge framework: documented, informal and tacit; the 40/35/25 percentages are examples

The diagram shows the vendor framework, not measured shares of your company’s knowledge. The categories can overlap.

The documented layer is what Tallyfy has been doing for the last 11 years. Workflow steps, approval gates, conditional logic and version-controlled procedures make repeatable work easier to run. A stable rule helps: given this trigger, do that thing, escalate to this person, finish in this state. That does not make every documented task safe to automate. Tallyfy’s own tribal-knowledge piece discusses the cost of losing expertise. A more precise number comes from Panopto’s 2018 survey: 1,001 US workers at employers with at least 200 staff reported that 42% of their job knowledge was unique to them. That is a survey estimate of knowledge, not a test of how much work replacements could perform.

The informal layer is where I see a useful role for AI. Capture judgment, exceptions and edge cases that have stayed in someone’s head. Try a Claude project or a custom GPT against real questions and check whether it can find the right precedent. The popular pitch of “AI replaces everyone” is rubbish; the argument here is narrower. (June 2026 note: this band is no longer hypothetical. Anthropic’s March 2026 Economic Index reports that about 49% of jobs already had workers doing at least a quarter of their tasks with Claude, and that people six months in showed roughly 10% higher success rates. The middle band is where that usage is landing.)

(November 2026 clarification: the 49% is coverage of occupational task categories in observed Claude use, not the share of workers using Claude. The roughly 10% success difference is an association with account tenure, not proof that six months of use caused it. The report does not measure the 35% knowledge band in this framework.)

The tacit layer is Polanyi’s paradox territory. In The Tacit Dimension, published in 1966, Michael Polanyi wrote that “we can know more than we can tell.” Think of reading a tense conversation, judging a vendor’s assurance or a surgeon recognizing a pattern. These are examples of hard-to-explain judgment, not measured speed comparisons. Eduardo Levy Yeyati argues at Brookings that tacit knowledge limits generative AI substitution. That is a serious objection. It does not establish a permanent 25% boundary: David Autor’s account also discusses machines learning from examples without someone spelling out every rule.

That is the scope cut. This post isn’t claiming total AI replacement of knowledge workers. The argument concerns the recall-heavy parts of work. Use workflow software for stable procedures, assign captured recall work to AI after checking its performance and keep people responsible for relationships and difficult judgment. It is the same reason AI does tasks, not jobs: a job is a bundle, and only some of the strands in the bundle are recall.

Why pay attention to mid-size companies?

Consider a company with 50 to 500 employees and little backup for key roles. A larger enterprise may have several people who know a system, though size alone does not guarantee that. A sole proprietor also has knowledge at risk, but the handoff problem is different. My concern here is the mid-size team with enough complexity to need specialists and too little cover when one leaves.

Here is an illustration, not a workforce benchmark. Suppose a 200-person operations team has 8 to 10 senior operators carrying informal knowledge. Assume annual turnover of 15% and a 6-to-12-month learning period for a replacement. Their knowledge includes vendor quirks, customer history and “we tried that in 2021 and here is what happened.” If none of it is recorded, each departure creates work the team has already done. Replace those assumptions with your own figures before estimating the cost.

That’s the risk to measure.

In advisory work with mid-size companies, this shows up the same way every time. The CEO knows there’s “tribal knowledge” but can’t name it. The HR director knows turnover is expensive but quotes the recruiter fee and the productivity ramp, not the forgetting cost. The forgetting cost is the bigger number. It’s just invisible, because nobody measures relearning. The post-hire onboarding kludge, 12 months of relearning repeated across 1 or 2 departures per year across a 30-person operations function, is what compounds the cost. Mid-size operators don’t have the recruiter overhead an enterprise carries to absorb this loss, and they don’t have the founder-attention slack a 10-person startup uses to paper over it. They sit in the band where the curve bites and the slack is gone. That is why I keep landing on the same recommendation: design the role around the substrate that doesn’t forget, and put the human on the parts the substrate can’t do.

The advisory figures above aren’t population estimates for all mid-size businesses. A maintained record can persist without biological forgetting, but AI retrieval and answers still need checking. Keeping a record doesn’t guarantee that a model will use it correctly.

Maybe I’m wrong here. But every consulting engagement I run lands in the same place, and the pattern is what convinced me to write Tallyfy in the first place.

Daniel Miessler argues that AI will replace knowledge work as the “articulation gap” closes: people explain expertise and AI accumulates it. His argument is spot on about replacing the recall-heavy parts of knowledge work. That’s my judgment; the forgetting curve doesn’t prove it. His claim that AI keeps it forever is a thesis, not a guarantee about a particular product. The case still depends on what gets stored, what remains accessible and whether the system retrieves it accurately.

What this changes about the next hire

Three questions to ask before opening a new requisition. Spend ten minutes on a first pass, then test the assumptions before changing the posting.

First: what fraction of the role is retention-critical? Meaning, the value depends on remembering rules, precedents, recent decisions and edge cases. If your estimate is above 70%, I would review the role for AI support before hiring. That is a rule of thumb, not a research threshold. Test accuracy, review time and the consequences of a missed exception.

Second: what fraction of the role is relational? Meaning, the value depends on empathy, persuasion, negotiation and sustaining trust. If your estimate is above 50%, my starting point is to hire a person and use AI to help them. Again, that is a judgment call, not a measured dividing line.

Third: what fraction of the role depends on judgment the practitioner struggles to articulate? This is the Polanyi problem. A job description or uploaded manual may capture very little of that skill. Do not assume AI can take it over, or that it never could. Look for evidence on the actual task and allow time for people to learn from experienced colleagues.

Hypothetical worked example. You’re replacing a junior contract reviewer who handles 60 SaaS renewals and vendor agreements a quarter. Q1: retention-critical, maybe 85%. Q2: relational, maybe 5% (most of the work is reading documents, not negotiating in a room). Q3: tacit, maybe 10%. I would test an AI workflow with a qualified human at the verification step before filling the role unchanged. Those percentages alone do not establish that it will be accurate or cheaper.

Different hypothetical example. You’re replacing a head of customer success at a 250-person SaaS company. Q1: retention-critical, maybe 25%. Q2: relational, maybe 65% (the job is reading customer emotional signals, retaining executive relationships, knowing when to escalate). Q3: tacit, maybe 10%. This is the wrong role to AI-replace. AI here is a tool the head of CS uses, not the head of CS. The actual work is sitting in a room with a frustrated VP of revenue who has just lost an executive sponsor at his largest customer, reading the body language of his second-most-senior account manager, and deciding whether to escalate to the CEO right now or after dinner. I would keep a person responsible for that conversation and the decision. A Claude project could help them prepare, but I would not hand it the relationship. The pattern that keeps showing up across hiring conversations is teams over-applying the AI substitution to roles where the value is relational, and under-applying it to roles where the value is recall.

The three-question screen does not tell you to hire fewer humans. It tells you which roles to design around AI and which to design around a person. Same headcount question, more precise answer.

Limits and counter-arguments worth naming

Four objections matter.

The first: AI also forgets. A long conversation can exceed the available context, and a model can miss material that is still inside it. Lost in the Middle demonstrated retrieval failures in the models it tested. Catastrophic forgetting during fine-tuning is another problem: learning new tasks can degrade earlier capabilities. Sean Warman’s discussion raises the practical concern. Larger windows help with capacity: Claude reached 100K in May 2023 and 200K in November 2023. In November 2026, Anthropic’s current model table lists 1M-token windows. Capacity is not reliable recall. Human memory also has no single universal seven-plus-or-minus-two limit; Cowan’s 2001 analysis argued for about four chunks under more restrictive conditions.

The second: spaced repetition. It helps. Cepeda and colleagues’ review covered 317 experiments and found that spacing and the time until a later test jointly affect retention. The useful schedule varies. My concern for busy operators is the work of maintaining that practice. A daily Anki deck of vendor details is one possible approach, but training, reminders and access to current instructions can also support recall. The comparison should include those options.

The third: Polanyi’s paradox is the strongest objection. A surgeon’s pattern recognition or a senior buyer’s judgment can be hard to explain and transfer. Brookings makes a useful argument about that limit, but it does not measure a 25% band that AI can never enter. My case remains narrower: replace the recall-heavy parts of knowledge work with AI, while keeping people accountable for relationships and difficult judgment. That is my staffing recommendation, not a conclusion proved by the forgetting curve. A system still has to show it can retrieve and use the relevant knowledge reliably.

The fourth objection worth naming briefly: the “AI is overhyped” reaction. I get it. Cycle after cycle of vendor pitches. Plenty of failed pilots. The forgetting research does not prove that AI should replace a particular role, and neither does a large context window. My narrower recommendation is to have AI do recall-heavy tasks once it has proved reliable on the work, and design the human role around the remaining responsibilities. The pattern of self-improving processes and Claude project knowledge bases gives you places to start. The companion question of how to prompt Claude to do this kind of cross-domain analysis is covered in the workflow-versus-persona prompt argument: describe the work, not the worker.

None of this means hire fewer humans. My recommendation is to hire people for the relationships and judgment the work needs, and stop hiring people for recall tasks that an AI workflow can perform reliably with appropriate checks. The next hire is correct. The next hire’s job description is what needs to change.

About the Author

Amit Kothari is an experienced consultant, advisor, coach, and educator specializing in AI and operations for executives and their companies. With 20+ years of experience, he is the Co-founder and CEO of Tallyfy® (raised $3.6m, the Workflow Made Easy® platform) and founding partner at Blue Sheen, which helps small and medium-sized companies get their whole team using AI. Originally British and now based in St. Louis, MO, Amit combines deep technical expertise with real-world business understanding. Read Amit's full bio →

Disclaimer: The content in this article represents personal opinions based on extensive research and practical experience. While every effort has been made to ensure accuracy through data analysis and source verification, this should not be considered professional advice. Always consult with qualified professionals for decisions specific to your situation.

Contact meMore about me

Blue Sheen helps small and medium-sized companies get their whole team using AI.

Contact me Follow 10k+