A five-part limited series on memory science and agent memory design
The Seams Are The Point is a five-part limited series adapted from a research document on memory science and the design of memory for AI agents — "Working Memory: What Biology Knows That Agent Memory Doesn't." Hosts Nora and Desmond trace what human memory research actually says, from Ebbinghaus's forgetting curve to the boundary between what a model may propose and what it must never author alone, while a third voice, The Record, reads the corrections and caveats the popular retellings drop. It is not a hype podcast about neuroscience-inspired AI — it is a forensic reading of a literature, with its own seams left visible: claims that failed fact-checking are recorded, not buried.
Ebbinghaus's famous forgetting curve turns out to measure savings, not recall, and Miller's “magic number seven” was never a working-memory limit — the industry spent seventy years building on a coincidence he explicitly dismissed. Nora and Desmond trace what capacity research actually supports: filtering, not window size, and a form of expertise reversal that turns a helpful default into a harmful one.
Every mechanism in human memory that looks like deletion turns out, on inspection, to be a change in accessibility. Nora and Desmond work through decay theory's collapse, the Bjorks' storage-versus-retrieval split, and why extinguished memories come back — landing on a design rule: retract by replacing, never by nulling.
Catastrophic interference wasn't networks being forgetful — it was a training regimen — and the classic two-systems argument turns out to be about representational format, not retention duration. Desmond and Nora unpack Tse's one-trial-learning rats to find the actual gate on what gets promoted to durable memory: consistency with what's already known, not recency or confidence.
More agents don't automatically remember more, and a mismatched pair performs worse than the sharper one working alone. Nora and Desmond survey transactive memory, hidden profiles, and Bahrami's asymmetric propose-and-dispose protocol to find out when multiple agents genuinely help — and when they just launder a single point of failure through committee.
The finale turns the show's own instruments on itself: a constructed scenario — suppose an agent tells a customer a refund went through — produces ghost memory, unjustified assertions, and a check that passes on a lie. Nora and Desmond derive where a model may only propose versus where it must never be the sole author — identity, state, and ordering — and lay out the measurement layers and kill criterion for what ships next.