In this paper, I argue that recent research on episodic memory supports a limited defense of the phenomena that Daniel Wegner has termed transactive memory. Building on psychological and neurological research, targeting both individual and shared memory, I argue that individuals can collaboratively work to construct shared episodic memories. In some cases, this yields memories that are distributed across multiple individuals instead of being housed in individual brains.
Central argument: Huebner defends a limited version of Wegner’s transactive memory thesis by grounding it in a constructive (rather than archival) theory of memory. He argues that some couples genuinely constitute distributed memory systems—but denies that Internet use does.
The Move: Wegner’s original framework treated memory as storage-and-retrieval: partners specialize, store different information, and retrieve it from each other via metamemories (like checking an index). Huebner shows this “archival” picture is empirically outdated. Memory is reconstructive—we store skeletal traces and flesh them out via counterfactual construction using schemas, environmental cues, and social input.
1. Situated Memories
Herbert Simon (1996) observed ants used environment for informational cues rather than representing it internally.
Evelyn Tribble (2011) discussed how early modern actors memorized roles in five shows per week using stage cues and interactions with other actors.
- But an internalist can absorb Tribble’s case: all cognitive work remains internal, dynamically updated against environmental information. This motivates Huebner’s higher bar.
Huebner adopts a traditional computational/representationalist framework (Weiskopf 2010) where cognitives systems are individuated by: 1) representational vocabulary, 2) operations on those representations, 3) control structures governing activation/inhibition.
- A cognitive system is whatever integrated network of mechanisms lies between transducers (input) and effectors (output).
The bar for extended cognition: external resources must be used within the computational architecture, not feed into it from the outside.
- Huebner will argue collaborative memory reconstruction (but not Internet use) clears this bar.
2. Transactive Memories
Huebner exposits Wegner’s empirical program and theoretical architecture for transactive memory systems (TMS).
Wegner et al. 1991: Romantic couples and ad hoc pairs memorize categorized item lists, both with and without assigned categories. Results:
- Couples with no guidance: best performance, M=31.4
- Ad hoc pairs with assigned categories: decent, M=30.14
- Ad hoc pairs, no guidance: worse, M=27.64
- Couples with assigned categories: performed worst, M=23.75
- The interference effect: imposing a different organizational scheme to encode information conflicts with and degrades familiar encoding.
- This memory degradation occurred at the group level, not the individual level, suggesting the couples developed a tacit system for dividing cognitive labor.
- Huebner: evidence for something beyond individual cognition—a transactive memory system.
Wegner posits a theoretical architecture of a computer network model. Partners maintain metamemory pointers indicating where information is stored—endogenously or exogenously. When a partner is unavailable, you get a feeling-of-knowing with no retrievable content.
- Huebner: this is a plausible computational sketch of an extended cognitive system. But it’s lacking mechanistic detail needed to distinguish genuine cognitive integration from mere informational dependence.
3. Googling Transactive Memory
Huebner presents Wegner’s attempt to extend transactive memory to Internet use, then explains why it fails
Sparrow, Liu, Wegner (2011) argue the Internet has come to function as an external memory, sometimes serving as part of a distributed cognitive system.
- People store metamemories, designating computers as memory storage locations and reflexively referencing them when presented difficult questions.
- People are less likely to encode information they think they can reference digitally.
Wegner and Ward (2013): the “Internet has become the external hard drive for our memories.” Wegner (2012) in the NYT: “We have all become a great cybermind.”
This extension only works if you accept an archival model of memory—information as inert packed stored somewhere and retrieved via search.
- But why should we accept that? The data equally support the banal interpretation: we use the internet as a tool without it becoming part of our cognitive architecture.
4. Constructing Memories
Huebner replaces the archival model of memory with a constructive one, which will ground his revised account of transactive memory.
Huebner rejects the archival model: memory as storage-and-retrieval.
- You encode a representation, file it away, and retrieve it later.
- On this account, extended memory would be easily implemented and Internet retrieval would count as remembering.
Memory is reconstruction, not reproduction. We don’t store complete records. We store skeletal representations—the gist. - This is significant because it will provide a more promising strategy for a mechanistic account of transactive memory.
- From a biological perspective, there’s little value to a robust, literal memory since the most important features of the environment are likely to change before we need or want to examine them again (Bartlett 1932).
- Memory is more like archaeology than archiving (Neisser 1967): “out of a few stored bone chips, we remember a dinosaur.”
Three empirical findings converging that memory is constructive:
- Construction from schema: gaps filled with culturally available assumptions.
- Students recalled Native American folktales years later and replaced the surreal aspects with culturally familiar assumptions (Bartlett 1932).
- Construction from repeated structure: you reconstruct thematic patterns across events rather than replaying specific ones.
- John Dean’s Watergate testimony wasn’t accurate in detail but captured thematic gist across many conversations (Neisser 1981).
- Construction is ongoing and permeable to input: you integrate post-hoc information into your “memory.’
- Post-event misinformation alters eyewitness testimony; people can be made to “remember” fabricated childhood events (Loftus 1974, 1995).
Remembering the past, imagining the future, and counterfactual thinking all share the same cortical network
I contend that a constructive theory of memory provides a more plausible foundation for transactive memory research of the sort that Wegner was interested in.
5. Transactive Memory Reconstructed
Huebner’s central positive argument. He combines the constructive theory of memory with Wegner’s transactive memory framework to produce a more defensible account of genuinely distributed memory.
Storage is individual; construction can be collective.
- Memory traces (skeletal encodings) are always stored in individual brains. But the constructive process—inferentially fleshing those traces into full episodic memories—can be distributed across people.
I suggest that the representations we store as memory traces are housed in individual brains, while the representations that we produce during the process of collective remembering are often realized by shared inferential processes that cross the boundary between brain, body, and world.
Wegner is right to claim that romantic partners and long-time friends sometimes become cognitively interdependent.
Long-term couples and friends collaboratively reconstruct shared memories through dynamic cuing and re-cuing. The counterfactual modeling that is typically internal gets externalized through communication.
Empirical support, Harris et al. 2011:
- Couples using dynamic, interactive cuing produce more detailed, episodic recall.
- Collaborative reconstructions allow people to recall details that neither partner remembered individually.
I argue that a couple should be seen as a transactive memory system when the processes required for episodic remembering are distributed between the members of a couple in this way. Significant computations can occur in a linguistic medium, at the interface between individuals.
Because counterfactual modeling operates over linguistically encoded representations, you can run it in natural language and therefore externalize it via language and gesture. When you offer a cue to a partner, you’re performing part of the constructive computation in a public medium. The inferential work can happen between people in conversation. This is what makes the dyadic interaction genuinely cognitive and not mere information exchange—the interface is doing constructive work.
- While each individual is also constructing their own internal representation, but it does not explain the behavior of the couple as a whole.
- The output, the reconstructed memory, is a product of the distributed process, not a sum of two individual, internal processes.
Three ways the transactive memory system breaks down:
- Incompatible expertise / authority claiming: one partner monologues.
- Repeated correction or passive acknowledgment (“uh-huh”): shuts down cuing dynamic.
- Retrieval-induced forgetting (architectural): one partner can block the other’s retrieval.
6. Wegner Reconstructed?
Huebner applies his reconstructed framework to adjudicate which cases count as genuine transactive memory and which don’t.
Wegner’s 1991 data showed functional specialization and group-level interference. But these coarse-grained results can’t distinguish between:
- couples who genuinely co-constructed memories through dynamic transactions
- couples who merely exploited each other as external storage
Why groups often perform worse than individuals:
- Most lab groups never develop the cross-cuing dynamics needed for collaborative construction. They default to exploitation—assuming what others will remember—and their strategies conflict.
- Only couples/friends who achieve genuinely collaborative construction outperform, producing richer episodic memories than either individual could alone.
The Internet does not act as an extension of memory because information flow is unidirectional and exploitative. Google doesn’t dynamically cue you back.
A sufficiently advanced AI could in principle constitute a TMS with a person if it could dynamically cue and re-cue skeletal representations, engage in shared counterfactual modeling, and jointly produce episodic or narrative representations.