2026/04 - Focus Week Recap on Temporary Models in the Hippocampus

Video Discussions
1

@jhawkins explores how the hippocampus supports rapid learning and temporary memory. Covering brain anatomy, grid cells, and insights from Henry Molaison (H.M.), the discussion examines how the brain forms episodic memories, learns spatial structure, and interacts with the cortex. It also challenges traditional hierarchical views, highlighting parallel processing, compositional learning, and the idea of the hippocampus as a “scratch pad” for real-time experiences.

Summary Video

Main Video

0:00 ​Introduction
1:01 Hippocampal Complex - Anatomy
4:31 Connectivity Between Regions in the Cortex
6:22 Patient H.M.
8:39 How does Hippocampus Learn Rapidly?
10:51 Grid Cells
13:52 Questions Around the Anatomy
17:16 Introspective Observations About Learning
22:24 Paper: Hippocampal Neurons Co-fluctuate in Freely Moving rats During Spatial Behavior
29:35 Learning First Happens in the Hippocampus
47:06 “Forking” Models
1:06:31 Questions
1:13:49 The Role of Context in Forking
1:28:21 Further Questions & Discussion
1:38:28 Follow-up Discussion: Can an Object ID “seed” a Reference Frame?

1 Like
2

Personal anecdote;

In 2024, my dad suffered an episode of transient global amnesia. Affecting 5 to 10 people per 100,000 each year, this rare condition involves a temporary one-day shutdown of CA1 pyramidal neurons in the hippocampus due to metabolic stress. Those neurons are the final output of the hippocampus, so during TGA, the hippocampus is completely disconnected from the cortex.

He had suddenly forgotten about COVID, the death of his father, my sister having gone to and graduated from dental school, and so on. I kept asking him questions to poke the extent of his memory loss. Roughly speaking, I estimate it spanned the 5 prior years.

He was only able to recall things that happened within the last minute, or more than 5-ish years ago. This indicates a very strong disruption of early long-term potentiation, with his one-minute memory likely being enabled by the electrical “echoes” of his neural spikes circulating across the cortex until signal attenuation.

His cognitive functions appeared otherwise normal. Similarly to Patient H.M., he was fully conscious and aware of his immediate surroundings, and had no trouble seeing, hearing, speaking, eating, moving his upper body, doing basic math, and answering abstract questions. To me, this signals that the hippocampus is probably more of a “sidecar” to the neocortex, rather than above it.

He was definitely able to recognize that he was in a hospital room upon seeing the medical equipment, so his object recognition capabilities were unaffected. His most striking behavior was asking “So, tell me, how did I get here?” hundreds of times, every minute, like a broken record. We told him, then he would immediately ask again. It felt eerily similar to talking with GPT-3; very short context and data cutoff a few years back. Fortunately, he made a full recovery within a few days, having however zero recollection of what happened during the episode.

I don’t know if this brings any new info to the table, but it’s quite an unusual experience that is oddly relevant to the topic, so I thought I’d share. I’m happy to answer more targeted questions if there are any.

Unfortunately, I was not acquainted with Thousand Brains at the time, so I did not perform further experimentation on the patient. However, a recent study highlighted a long-term 11.2% chance of a recurrent episode in TGA patients, so I am prepared if it rehappens… :face_with_monocle:

4 Likes
3

Wow, that is very interesting! Thank you for sharing! I might come back to you with more questions but here is one:

Did you get an impression of his ability to perform novel spatial tasks? I assume he had trouble learning the layout of the hospital and navigating to the place? I’m not sure if there were any examples of smaller-scale spatial tasks, like interacting with a new object or rearranging objects on a table. No worries at all if nothing comes to mind.

4

This is a great video. I loved the discussion on hippocampus. To me, it feels like the cortex is extracting patterns from streams of signals, discretizing them, remembering commonly occurring sequences in pattern neurons, and re-activating them when they happen again. That’s used to predict and decide short-term (system 1). And HCC is capturing all the active patterns in moment neurons. When pattern neurons decay, they forget parts of the sequences, only remembering the core parts that repeat consistently. When moment neurons decay, they lose the connection to the pattern neurons, turning moments into class neurons. Moment neurons are used to predict and decide long-term (system 2). Let me know what you think.

1 Like
5

There wasn’t that much opportunity for novel stuff. Here’s a full account of the event;

On a late Saturday afternoon, he was doing some house cleaning on the second floor. My mom went to see him, and he said he felt disoriented and forgot what he was doing. She talked with him a bit and quickly realized something was wrong. She asked him to follow her to the first floor, and he went down the stairs on his own without trouble. I was in their basement at the time, packing boxes of my old stuff, since I was moving out of the city a few weeks later.

She came down to see me in a state of distress, saying he lost his memory. I checked on him and questioned him a bit, eventually asking “How is grandpa doing?” (he passed away 3 years earlier) and my dad replied that grandpa was “doing well”… That was the “oh s**t” moment. We thought he might be having a stroke, so I sat him down while we looked for his medical insurance card. He said not to worry, that he was probably just tired… We found his card a minute later, and my mom drove him to the hospital, which was just 5 minutes away.

She demanded I stay home to keep an eye on her new exotic-breed kitten. For whatever forsaken reason, I somehow agreed and let her leave alone with him, but it quickly dawned upon me that my dad might be living his final moments. I phoned my sister and went to the hospital. When I arrived, he was already going thru CT scan. He was transferred to a temporary bed while awaiting results. He kept repeating “How did I get here?” like a broken record the whole time. We kept telling him and questioning him.

The scan came back clean; no stroke. I asked GPT and it replied it might be TGA, which the nurse had not heard of. He was then moved on a gurney to an ER bed. We kept talking with him. My mom got tired of repeating over and over, so she recorded a video of her response on her iPad, gave the tablet to him, and told him to play the video. He appeared able to interact just fine with the iPad, like he did every day back home.

It was getting past dinner time, so we got him a sandwich and a yogurt cup, along with some grapes I brought from home. He ate everything fine, I didn’t notice anything abnormal, apart from his constant questions about what was going on.

I unfortunately had a mild cold and I was starting to choke on mucus behind my mask, and he seemed stable, so I returned home. My mom stayed with him until he fell asleep, then came home to sleep, and went back the next morning. They kept him overnight on anticoagulants until an MRI scan in the morning, in case it might be a “mini-stroke” that the CT scan couldn’t pick up.

MRI came back clean, and his memorization had mostly recovered, so the doctor agreed with a TGA diagnosis, and he was discharged at noon. He had no recollection of what happened between onset and waking up, but remembered everything before and after.

My mom was by his side from onset to sleep, so I’ll check with her to see if she noticed anything unusual or novel. I really wish I had stayed longer and picked his brain more, but with a cold and all, I had to keep contact down to a minimum.

2 Likes
6

SILENT SYNAPES – SOME POINTS OF VIEW, AND POTENTIAL IMPLICATIONS

1. INTRODUCTION

Jeff Hawkins mentioned “silent synapses” in this video. I was fascinated, so I started to learn about them, then tried to think about possible implications.

In the learning phase, I learned that much more is known about them in rodent than in human brains, and in the hippocampus than in other areas. I also learned that AMPA receptors play a key role in un-silencing them then silencing them again.

What follows is constrained by what’s known scientifically, but I also do some speculation.

2. VIEWING SILENT SYNAPSES FROM VARIOUS POV’S

In the thinking phase, first I tried to inhabit different points of views:

- If I inhabit a postsynaptic neuron then silent synapses increase my options, and their noises are my danger. Many could influence me, but only some do, the rest are my future possibilities.

- As a presynaptic neuron, silent synapses increase my options but not my danger. With them, I tentatively bet on future relationships, most will never matter, but some may.

- IF EVER silent synapses are proven to lurk in some adult human cortical columns, then for me as a column, my learning can partly involve selection rather than purely construction.

3. IMPLICATIONS

Then I tried to understand what implications may arise. The first are already said in the above video: Silent synapses help explain being able to rapidly learn, to juggle flexibility with stability, and to have both stable structures and fleeting ones.

Another possible implication is to help explain some macro cognitive phenomena. For example, sometimes a mental experience seems to jump to an apparently unrelated one.

If Monty were ever to incorporate an analogue of silent synapses, the first implication above would be the immediate impact. But a deeper impact might be the use of the computational principle of using latent structure for prediction, and mechanistic explanations for some cognitive phenomena.

There are costs, however. While it would be cheap to pre-provision a silent synapse, once it has been unsilenced, it incurs costs just like any other.

Most importantly, a cacophony of noises would destabilise the TBT brain unless we design algorithms to generate, prioritise, silence, and prune silent synapses. The process of designing such algorithms – or having the algorithms reveal themselves to us from some fundamental rules – likely would teach us a lot, because the problem is general and powerful: Which latent structures should be active, which dormant, which to remove?

1 Like