# Your Brain Is a Sculptor, Not a Scribe *And every AI agent built today has the forgetting problem backwards* --- There's an old question philosophers have argued about for two and a half thousand years. When you were born, was your mind a blank page that the world has been writing on ever since? Or did you arrive already full of something — instincts, structures, possibilities — that life has been carving away at? Aristotle voted blank page. Locke agreed and named it: *tabula rasa*, the unscribed tablet. Leibniz disagreed — he thought the mind came pre-veined like marble, and experience was the sculptor working with grain that was already there. Kant tried to split the difference. Behaviorists picked one side. Chomsky picked the other. The argument ran for centuries because nobody could open a developing brain and just *look*. Last month, somebody finally looked. A team at the Institute of Science and Technology Austria, led by Peter Jonas, watched what happens to the memory circuits of baby mice as they grow up. They expected to see a brain filling itself in — more connections, more wiring, more structure as experience accumulates. They saw the opposite. --- ## The carving Here's what they did, in plain terms. The hippocampus is a small curled structure deep in the brain. It's the reason you remember anything at all — when surgeons removed it from a patient named Henry Molaison in 1953 to control his seizures, he lost the ability to form new memories for the rest of his life. Inside the hippocampus there's a region called CA3, special because its neurons talk to each other in a giant looping conversation. That loop is where memories actually get stored. The Austrian team listened in on those conversations at three ages: newborn mice, teenage mice, adult mice. They tested nearly eight thousand possible connections between neurons. The newborn brain was *crowded*. Roughly one in sixteen neurons was wired to any given other neuron. Connections were dense, local, and largely random — like a city where everyone knows their neighbors but nobody knows anyone across town. The adult brain was *sparser*. Only about one in fifty neurons stayed connected. But the connections that remained weren't random anymore. They formed specific patterns, reaching across long distances to specific partners. The city had grown up. People had figured out who was actually worth talking to. The brain wasn't filling itself in. It was *carving itself out*. We are not born blank. We are born **overstuffed**. And growing up is mostly the process of losing connections — keeping the ones that matter, dropping the ones that don't, until what's left is the specific person you become. There was a second finding, too. In newborn mice, a single neuron firing was often enough to make its partner fire. The young brain runs hot — almost anything can set off almost anything else. In adult mice, you needed about twenty-five separate inputs all arriving at roughly the same time. The synapses had become weaker, more demanding, harder to trigger. This sounds like decline. It isn't. It's the difference between an alarm system that goes off when a leaf brushes the window and one that only goes off when there's actually a burglar. The young brain encodes everything because it doesn't know yet what matters. The adult brain encodes selectively because it's learned what to ignore. This is also why you probably can't remember anything from before you were three or four. The memories formed at the time. But the wiring that stored them got pruned away as your brain reorganized itself. The keys were thrown out. The rooms still exist somewhere, but you can't get in. --- ## Forgetting is not a bug This is the part I want to dwell on, because I think it's where the most important misunderstanding lives. For most of the history of AI, forgetting has been treated as a *defect*. The classical story, called catastrophic forgetting, was first documented in 1989: train a neural network on Task A, then train it on Task B, and watch it lose almost everything it knew about Task A within a few epochs. The weights that encoded A get overwritten by gradient descent reaching for B. The network doesn't gracefully prune. It just *forgets violently*, in ways no biological system would tolerate. Most of the modern continual learning literature is about preventing this. Elastic Weight Consolidation, published in 2017, identifies which weights were "important" for previous tasks and adds a penalty that resists changing them. Replay buffers shuffle old examples back into training to keep them fresh. Spaced repetition schemes copy the Ebbinghaus forgetting curve from 1885 — review at one day, then two, four, seven, fifteen — and use it to schedule what gets re-shown to the model. Notice what's happening here: in all of this work, the goal is to *fight* forgetting. Forgetting is the enemy. If we could just hold onto everything, the assumption goes, we'd have a system that learns continuously like a human. But humans don't learn continuously by holding onto everything. Humans learn continuously by **forgetting on purpose**. The hippocampus paper is, in part, a careful empirical demonstration of exactly this. Roughly two-thirds of the connections in the newborn CA3 network are *deliberately destroyed* over the course of normal development. Not because the brain ran out of space. Because forgetting *is what makes the remaining structure mean something*. This is the inversion the field has missed. Forgetting isn't a bug in biological memory that we need to engineer around. Forgetting is the mechanism by which biological memory acquires its shape. --- ## Why current agent memory gets it backwards The other place AI talks about forgetting is at the agent layer. A wave of frameworks now gives LLM-based agents persistent memory. The standard pattern, as of 2026, looks like this: when an agent has a conversation, a smaller model extracts "facts" or "preferences," writes them to a vector database, and on future invocations retrieves the relevant ones by similarity search and injects them into the prompt as text. *"User prefers concise responses. User works at Acme. User has two cats named Luna and Mochi."* These systems usually do implement some kind of forgetting. Almost all of them use some variant of three mechanisms: **Time-to-live tags.** Each memory gets an expiration date based on its category. Allergies are immutable. Casual preferences expire in 90 days. The system runs a sweeper. **Decay functions.** Often borrowed directly from Ebbinghaus, sometimes ACT-R from cognitive science. Memory relevance scores decay exponentially unless something refreshes them. **Refresh-on-read.** When a memory gets used, its decay clock resets and its score boosts. Used memories survive; unused ones drift toward expiration. This is biologically inspired in the most literal, surface-level sense. It's also a fundamentally different thing from what the hippocampus actually does, and the difference is the whole point. Ebbinghaus decay is *temporal*. The system forgets things because they got old. The clock ran out. There was no review. Hippocampal pruning is *structural*. The system forgets things because they didn't earn their place in the network. They didn't participate in the patterns that proved useful. They were redundant with stronger pathways. They were locally clustered when the system needed long-range connectivity. The pruning happens because of *what the connection does*, not because of *when it was formed*. These produce wildly different behaviors. A TTL system will forget the brilliant, rarely-mentioned insight a user gave it three months ago — because it's old, even though it's central. It will keep the trivia about cat names indefinitely, because the user re-mentioned it last week. The clock has no taste. A decay-plus-refresh system has the opposite problem. It accumulates whatever gets *re-used*, regardless of whether re-use was a sign of importance or just a sign that something kept showing up. Frequently-mentioned trivia outlives rarely-mentioned wisdom. The recency boost is a popularity contest, and popularity is not the same thing as structural value. And neither of these systems does the thing that actually matters in the hippocampus, which is to **let the topology of the connections themselves decide what survives**. CA3 doesn't decay memories on a clock. It carves a network. Connections that participate in useful patterns get strengthened. Connections that don't get pruned. The substrate itself is the judge. You can see why current architectures can't do this. There is no substrate. There is a database of independent records. The records have no relationships to each other beyond what someone explicitly extracted. There is no graph for pruning to act on. There is only a list, and lists can only forget by aging or popularity, because those are the only signals a list contains. --- ## What tabula plena would actually require If you took the hippocampus seriously, here's what an agent memory system would have to look like: It would have to **start full**. Not empty. The first time an agent enters a domain, the substrate shouldn't be a clean slate that accumulates facts as they arrive. It should already be richly over-connected — full of hypothesized relationships, candidate patterns, possible interpretations of what it might encounter. Most of these will be wrong or redundant. That's fine. That's the point. The young brain is mostly wrong. It's wrong in the *right way* — wrong with enough variety that experience can carve toward correctness. It would have to **store patterns, not facts**. The unit of memory in CA3 isn't a record. It's a sub-network — a specific configuration of which cells fire together when a pattern reactivates. The agent equivalent isn't *"user prefers concise responses."* It's the recurring shape of how interactions with this user actually go: which contexts trigger which kinds of follow-up, which intermediate states preceded successful resolutions, which sub-graphs of behavior repeat across sessions. The pattern is the memory. A description of the pattern is just a description. It would have to **prune by structural value, not by clock**. The right question is not "how long has this memory been here." The right question is "is this memory part of any useful pattern, or is it isolated?" Connections that don't participate in motifs should be pruned. Connections that do should be reinforced. The judge is the topology, not the calendar. It would have to **shift its operating regime as it matures**. Early in a domain, the substrate should be jumpy, easily activated, willing to encode anything. Once the domain has been learned, the substrate should become demanding, requiring multiple converging signals before firing, biased toward selectivity over recall. This is the temporal-summation-to-spatial-summation switch the Austrian team found in CA3 cells. It is missing entirely from production agent systems, which apply the same retrieval rule on day one and day one thousand. It would have to **accept that some memories will be lost on purpose, and that this is good**. An adult mind can't recall most of its toddler experiences because the wiring that stored them was pruned. We don't consider this a failure of human memory. We consider it part of what makes adult memory *work* — the carving that turned a generic infant brain into a specific adult one. Agent memory systems treat any forgetting as a defect to be minimized. That bias is what keeps them shallow. --- ## The deeper lesson There's a pattern across the AI literature on forgetting, and it's worth stating plainly: every current approach treats forgetting as either *something to prevent* (continual learning, EWC, replay buffers) or *something to schedule* (decay curves, TTLs, refresh-on-read). Neither of these is what biology does. Biology treats forgetting as **the act of becoming specific**. You are not the sum of what you've learned. You are the residue of what you've forgotten on purpose. The hippocampus paper is, at its deepest, a quantitative demonstration of this fact: roughly two thirds of the connections present at birth are gone by adulthood, and the ones that remain are the ones that earned their place by participating in patterns that proved useful. The brain is a sculptor that spends its first decades carving a generic block of marble into the specific shape of you. If we want agents that can become specific the way you became specific — that can start with too much, lose what didn't matter, keep what did, and end up with a shape that's earned rather than accumulated — we need to stop treating forgetting as a bug to fix and start treating it as the primary mechanism by which memory acquires meaning. The mice figured this out forty million years ago. We're still catching up. --- *Based on Vargas-Barroso et al., "Developmental emergence of sparse and structured synaptic connectivity in the hippocampal CA3 memory circuit." Nature Communications, April 2026.*