You have probably had the realisation — I see exactly why I do this — and then done it again on Monday. The science of change explains why. Real change in people tends to come in jumps rather than a smooth slope; the jump is often preceded by a wobble, sometimes by a spell of feeling worse; and what decides whether it lasts may be what happens after. A German school of therapy has been measuring this, day by day, for twenty years. Between them sits something you can test.
Most people who have tried to change something know the pattern. You have the insight — in therapy, in a book, at 2 a.m. — and it feels like the problem is solved. Then life resumes, and so does the old behaviour.
This is not a character flaw. It is one of the best-measured facts in psychology. In 2006 Thomas Webb and Paschal Sheeran gathered 47 experiments where researchers first changed people's intentions and then watched what they actually did (Psychological Bulletin). A medium-to-large shift in intention (an effect size, d, of 0.66 — a standard way of saying "how big") produced only a small-to-medium shift in behaviour (d = 0.36). Roughly, about half of the change in wanting survived into doing. Insight is real. It just isn't enough.
So if understanding doesn't do it, what does? The answer from the last twenty-five years of research is stranger, and more hopeful, than "try harder."
The textbook picture of therapy is a gentle downhill slope: a bit better each week. In 1999 Tony Tang and Robert DeRubeis looked at session-by-session scores in cognitive therapy for depression and found something else: many patients improved in a single large drop between one session and the next — a sudden gain — and those patients did better in the end. In 2020 Jonathan Shalom and Idan Aderka pooled 50 studies (Clinical Psychology Review). Sudden gains predicted better outcomes at the end of treatment (g = 0.68) and at follow-up (g = 0.61), across different therapies, disorders and settings. Notably, the benefit was smaller where gains often reversed — keep that in mind.
Outside therapy, the same shape shows up in habits. Phillippa Lally's team at University College London followed people building a new daily habit (European Journal of Social Psychology, 2010). Automaticity — how much the behaviour runs without effort — climbed fast at first, then levelled off. The median time to the plateau was 66 days, with a huge range: from 18 days to a projected 254. A 2024 review by Ben Singh and colleagues (20 studies, 2,601 people, Healthcare) found medians of roughly two months and an individual range from 4 to 335 days; the biggest single factor was how complex the behaviour was. And a 2023 study in PNAS (Anastasia Buyalskaya, Katherine Milkman, Angela Duckworth, Colin Camerer and colleagues), using machine learning on tens of millions of real-world records, found gym habits took months to form and hospital hand-washing habits took weeks. The "21 days" figure you have seen everywhere is not supported by any of this.
There is a second piece of good news: people can change things that feel like fixed personality. Brent Roberts and colleagues reviewed 207 studies (Psychological Bulletin, 2017). Over an average of 24 weeks, interventions (mostly therapy) produced meaningful change in personality measures (d = 0.37), mainly in emotional stability, and the change held in follow-ups. The type of therapy mattered little.
Here is the finding that should change how you think about a bad patch. Adele Hayes and colleagues tracked 29 people in a therapy for depression that deliberately had them face what they had been avoiding (Journal of Consulting and Clinical Psychology, 2007). A substantial share of them had a temporary depression spike — a period of feeling noticeably worse — during that phase. And the spike predicted lower depression at the end, not higher. Their journals during the spike showed more emotional and cognitive "working through." The worsening looked like an old pattern coming loose.
Researchers who study systems that switch between states — lakes, ecosystems, the climate — have a name for this. Just before a system flips, it often becomes unstable: it swings more, and takes longer to settle after a knock. In 2020 Merlijn Olthof, Günter Schiepek and colleagues tested this in patients with mood disorders who filled in a short questionnaire every day during intensive therapy (Clinical Psychological Science). Periods of "critical fluctuation" — unusual turbulence in the daily scores — predicted a sudden shift within the next four days. Crucially, the shift could go either way: towards a sudden gain, or a sudden loss.
That last detail matters. Instability is not a promise of improvement. It is a sign that the system is loose — that change, in some direction, is possible.
"Most change is gradual; the jumps are noise." This is the serious objection. If you measure anything weekly with a questionnaire, random ups and downs will sometimes look like a "sudden gain," and a lucky low score will look like a breakthrough. Shalom and Aderka themselves found that studies using looser definitions of a sudden gain reported bigger effects than studies using the original strict criteria — exactly what you'd expect if some of the signal is measurement. The fair reading: sudden gains are real on average, but any one jump in any one person may be noise.
"The warning signs don't work for individuals." The early-warning idea is elegant, but the most careful individual-level test is sobering. In 2025 Arnout Smit and colleagues in Groningen tracked 37 people coming off antidepressants, five times a day for four months (Clinical Psychological Science). The textbook warning signs preceded a return of depression in only about a third (32.9%). The same group has published a pointed perspective questioning whether these signals can predict symptom changes from a person's emotion diary at all.
"Memory can be rewritten" — the flattering idea that failed its test. A popular account of deep change says that when an old emotional memory is recalled, it briefly becomes editable, and new experience in that window can rewrite it (Lane and colleagues set out the therapy version in Behavioral and Brain Sciences, 2015). In 2010 Daniela Schiller and colleagues reported in Nature that recalling a fear memory and then running extinction training inside that window stopped the fear returning. It is a lovely story — and in 2020 Anastasia Chalkia, Tom Beckers and colleagues in Leuven ran a large, preregistered direct replication (Cortex): no lasting benefit over ordinary extinction. Chalkia's team also reported irregularities in the original paper's exclusion criteria; Schiller's group stands by the findings and says most later replications succeeded. The idea remains alive in animal work. As a settled account of how people change, it is not there — hence the 2/5.
"It's just the relationship." A long line of researchers (Wampold most prominently) argue that what all good therapies share — a trusted relationship, a believable explanation, hope — does most of the work, and the specific technique matters less. Roberts' finding that therapy type barely affected personality change is consistent with that. This piece doesn't need to settle that fight; it is asking about the shape of change, not which technique produces it.
The English-language conversation about change is mostly about techniques and trials. A German-speaking tradition has spent decades asking a different question: what does change look like from the inside of a system, day by day?
It starts with Klaus Grawe, a Bern psychotherapy researcher whose 1994 book Psychotherapie im Wandel — von der Konfession zur Profession ("Psychotherapy in transition — from confession to profession", with Donati and Bernauer) reviewed some 900 controlled studies, and who went on to propose that all effective therapy works through a handful of shared change factors. Two of his German terms are worth learning. Ressourcenaktivierung — resource activation: switching on what is already working in a person. And Problemaktualisierung — problem actualisation: making the problem live in the room, felt rather than just talked about. Grawe's later book Neuropsychotherapie (2004; English 2007) framed suffering as inconsistency — a mismatch between what a person needs and what they experience — and therapy as the slow restoration of consistency.
His intellectual heirs took the next step. Günter Schiepek, working in Salzburg and Munich and drawing on the physicist Hermann Haken's "synergetics" (the study of how systems self-organise into new patterns), built the Synergetic Navigation System (SNS): patients rate a short, personalised questionnaire every day on a phone or computer, and the therapist and patient look at the resulting curves together. In a feasibility study in an inpatient and day clinic, 151 patients completed their daily ratings with a median compliance of about 89%. What the curves keep showing — and what the 2020 Olthof study formalised — is the pattern in §3: turbulence, then a new pattern. In 2014 Schiepek, Igor Tominschek and Stephan Heinzel tested the synergetic model directly on daily data from inpatient therapy (Frontiers in Psychology). They found discontinuous jumps rather than smooth gains, critical instability just before the jumps, and — the part the English literature rarely mentions — that stable surroundings were a precondition: a steady therapeutic relationship and a safe setting were what allowed the destabilisation, and the re-stabilisation after it. (One of their hypotheses failed: motivation to change did not behave like the "dial" that drives the shift.) The therapist's job, in this picture, is not to push the person into a new shape, but to make the old one loose enough, and the setting safe enough, for a new one to form.
What the German account adds to the English one is the idea that instability is not the enemy of therapy — it is its working material. The English-language research largely treats a bad week as a setback to be minimised. The Grawe–Schiepek line treats it as the moment the system can actually reorganise.
Put the pieces side by side. Intentions alone barely move behaviour (§1). Real change comes in jumps (§2). Jumps tend to be preceded by turbulence, sometimes by feeling worse (§3). Sudden gains that reverse carry less benefit (§2). Habits take weeks to months of repetition in a stable context to settle (§2). None of these fields joins them up. Here is a way to:
Lasting change has two phases, and most attempts only do one. First, loosening: the old pattern has to become unstable — felt, activated, disrupted (Grawe's Problemaktualisierung, Hayes' spike). Second, locking: the new pattern has to be repeated in a stable setting until it holds on its own (Lally's plateau). Insight without locking fades by Monday. Repetition without loosening just digs the old groove deeper. The durability of a change depends less on how big the jump was than on whether the system settles down afterwards.
This is a proposition — my interpretation — not a finding. To be straight about novelty: the German synergetic school already describes destabilisation followed by re-stabilisation, so the two-phase picture is theirs. What I am adding is a narrower, checkable claim joining their work to the sudden-gains and habit literatures: that how a person settles after a jump — measurable as a fall in day-to-day variability — predicts whether the change lasts.
The build (simulation). A simple model of a person as a ball in a landscape with two valleys — an "old pattern" valley and a "new pattern" valley — nudged by daily random life events. Two dials: how deep the old valley is (how entrenched the habit), and how quickly the new valley deepens with each day spent in it (repetition-based consolidation). Run thousands of simulated people through a "therapy" phase that temporarily shallows the old valley (loosening).
The data arm (re-analysis). Daily-rating therapy datasets of the kind the Schiepek and Olthof groups have collected — accessed with permission and checked through this programme's feasibility gate before anything is announced.
Pre-registered predictions:
Why it's worth running. If P1 holds, it gives a practical, measurable marker for "this change is consolidating" — useful to anyone tracking their own progress or a client's. If it fails, the two-phase proposition is wrong in its distinctive part, and I'll say so. The simulation alone cannot prove anything about real people; it tests whether the idea is coherent and what effect sizes the data arm would need to detect, which is exactly the job the Optionality ageing study's simulation did first.
The hard line: nothing here is a reason to stay in a period of worsening without support, or to stop a treatment. If things are getting worse, talk to whoever is supporting you. Feeling worse is sometimes part of change; it is also sometimes just worse.
(David Fleming's own interpretive framework — a way of seeing, not established science. It models structure, not experience.)
Recursive Field Theory describes a stable pattern as a configuration that keeps re-closing on itself — each cycle reinforcing the next. In that language, a habit or an emotional pattern is a loop with closure; change means reducing closure enough for the loop to open, then letting a new loop close and deepen through repetition. The two-phase proposition is that picture, made checkable: §7 names the measurable trace (variability up, then down) that the framework would be wrong about if it didn't appear. Offered as interpretation.
Adversarial self-check before saving. Three things caught and corrected: (1) my first draft said the 2020 Olthof study showed instability "predicts breakthroughs" — it predicts sudden shifts in either direction, gains or losses; §3 now says so. (2) I had used memory reconsolidation as supporting evidence for the proposition; after checking the 2020 Leuven registered replication, it is now in §4 as contested and rated 2/5, and the proposition does not lean on it. (3) I nearly repeated the "21 days to form a habit" figure as a straw man attributed to Lally; her study never claimed it, and §2 now gives her actual numbers. A fourth check cut my own novelty claim down: Schiepek, Tominschek and Heinzel (2014) already describe destabilisation-then-restabilisation, so §6 now credits them with the picture and claims only the narrower settling-predicts-durability test. All statistics in this piece were checked against the source abstracts; Grawe's study count is given as "some 900" because I could verify the order of magnitude but not the exact figure from an English-language source. The proposition and the study design are mine and labelled as such. (Verification pass, 13 October 2026: Webb & Sheeran's d = 0.66 → 0.36, Lally's 66-day median and 18–254 range, Shalom & Aderka's g = 0.68 / 0.61 across 50 studies, Roberts' d = 0.37 across 207 studies, and the Olthof and Schiller/Chalkia results were all confirmed against source. Two figures that could not be confirmed against a primary source were softened rather than risk a loose number: the exact participant total for the Shalom–Aderka meta-analysis, and the precise proportion of Hayes's 29 patients who showed a depression spike — the finding and its direction are unchanged.)
Format & time. Reflective structured learning · ~30–45 minutes.
Learning outcomes. After this you will be able to: describe the evidence that therapeutic change is often non-linear (sudden gains), and the main objection to it; explain the intention–behaviour gap with its effect sizes; summarise the evidence on habit-formation timelines; describe the contested status of memory reconsolidation in humans; and state one falsifiable, testable version of "lasting change requires loosening, then locking."
Reflection prompts (write 3–5 lines each, keep for your CPD record):
CPD-eligible structured learning; self-certified, not statutory-regulator endorsement. Log only genuine time spent. Not medical or psychological advice.
Primary sources linked in the published version. The weekly Investigation is free, in full — including the disputes. Paid members get the monthly Cross-Domain Proposition, where a testable idea like §7 is developed and taken toward a registered study. Recursive Field Theory is the original work of David Fleming. Not medical advice.
Join the Frontier. The weekly Investigation is free, in full, every week.