Transactive memory is one of the older ideas in cognitive science, and one of the most useful for thinking about AI. Long before search engines, long before AI, Daniel Wegner noticed something obvious that no one had named: long-term couples remember things together. One partner is the keeper of birthdays, the other is the keeper of where the keys live. Neither feels the need to remember the other's domain, because the system as a whole always knows.
That framework — developed in the 1980s, extended to search engines in 2011, and now being extended to AI — is one of the cleanest ways to understand what changes when a powerful new memory partner enters daily life.
The original framework
Daniel Wegner introduced the term transactive memory in 1985. The core idea is that close, stable groups — couples, families, teams that have worked together for years — function as a single memory system distributed across members. Each member specialises. Each member knows what the others know. Crucially, each member also knows that the others know, and where to retrieve the information when needed.
The result is a system that, in aggregate, remembers far more than any individual member could alone. The cost is that any individual member, considered alone, has gaps that look like deficits but are actually specialisation. A spouse who can't remember a single birthday is not failing at memory; they are running on a transactive system where birthdays live in their partner's memory and the relevant retrieval pathway is "ask."
This is not a metaphor or a loose analogy. Wegner and colleagues went on to demonstrate measurable effects: long-term couples encode information differently when their partner is present, retrieve different items than they would alone, and show coordinated specialisation that increases over years together. The Wikipedia entry on the Google effect collects the broader research trajectory.
What Sparrow extended in 2011
In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner (the same Wegner) published "Google Effects on Memory" in Science — the foundational empirical paper for what is now called the Google effect. The conceptual move was simple and important: treat the search engine as a transactive memory partner.
The experiments tested whether participants encoded information differently when they expected to be able to look it up later. They did. When information was expected to remain accessible, participants remembered the information itself less reliably and remembered the location of the information better. The brain was doing exactly what it does in human transactive systems — specialising. The search engine was treated as the partner who remembers the facts; the user's memory specialised in the meta-knowledge of where to find them.
The 2024 Frontiers in Public Health meta-analysis consolidates the literature that built on Sparrow's framing. The Google effect is real in aggregate; the effect size depends on context. The underlying transactive-memory mechanism is the most defensible part of the framework.
AI as the next transactive partner
AI assistants slot into the transactive-memory frame naturally. They are an available memory and reasoning partner, accessible on demand, capable of remembering far more than any single human can. The encoding shift the Google effect identified for search engines should — and on early evidence does — extend to AI use, in a heavier dose.
There are two important differences worth thinking about clearly.
The first is asymmetry. Human transactive systems are mutual. Each partner contributes; each partner also remembers part of what the other will need. The system is stable because the load is shared. AI is not a mutual partner. You draw from it; you do not also support it. The transactive system is therefore one-sided — all the offloading goes in one direction. The framework predicts that the encoding shift on the human side will be larger than in mutual human partnerships, simply because there is no balancing flow.
The second is range. Search engines mostly offload factual retrieval. AI assistants offload synthesis, drafting, planning, decision support, even emotional processing in the case of users who use AI for conversation. The transactive surface is much wider. The 2025 MDPI paper on AI and cognitive offloading notes this directly. The transactive-memory framework would predict that as the AI partner's competence broadens, the user's specialisation narrows — possibly to the meta-skill of prompting and evaluating, possibly to something even thinner.
What this does not say
It is worth being clear about what the transactive-memory framework does not claim.
It does not say that using AI is bad. Transactive systems are how human groups have always remembered. Couples, teams, families, professional communities — all of them rely on distributed memory. The framework is descriptive, not prescriptive. It says: this is how memory systems involving multiple agents work.
It also does not say that AI use causes cognitive decline. The pattern is encoding specialisation, not loss of capacity. A spouse who never remembers birthdays is not bad at memory; they are running on a system where memory is allocated differently. By the same token, an AI user who never drafts the first sentence is not necessarily losing the ability to draft; they are running on a system where drafting is allocated differently. Whether the underlying drafting skill atrophies over time — which is a separate, more empirical question — depends on whether it is still practised somewhere.
What this means in practice
The transactive-memory framework points to a clear practical implication. If the system you are in is one-sided — all offloading in one direction, no balancing flow back — then the human side's specialisation will narrow over time. The way to keep a skill in the system is to keep practising it in the system. This is the standard offloading-literature advice in transactive language.
For daily life this looks like the same kind of small, regular, unaided practice the rest of the offloading literature recommends. Write the first paragraph yourself before showing it to AI. Estimate a number before asking. Draft a plan, then ask AI for feedback rather than asking AI to plan. The point is not to refuse the transactive partner. The point is to keep your share of the system non-trivial.
That is the bet Senwitt is built on: a small daily moment where the thinking is yours, across a mixed set of skills, so that the underlying capabilities continue to receive reps even when the rest of the day is partnered with AI. The research/transactive-memory-and-ai page goes into the full version of this argument.
The transactive-memory framework is forty years old. It still has the cleanest vocabulary we have for talking about what changes when a powerful new memory partner enters the room.
There is a final useful turn worth naming. When Wegner first described couples as a transactive system, the point was descriptive — this is how human groups remember, and have always remembered. The system is stable because the parties depend on each other and stay in proximity over years. When AI enters that frame, the system is still functional but the stability assumption no longer holds in the same way. You can leave a long-term partnership and slowly rebuild the personal store of what your partner used to hold. The transactive-memory literature has documented exactly that recovery curve. The same recovery is available for AI-offloaded skills, but only if the underlying skill was practised somewhere along the way. The practical implication is gentle: keep the human side of the transactive system non-trivial not because the AI partner will leave, but because the recovery path needs an unexercised skill to still be reachable. That is the entire bet.
