How memories are stored
How experiences are physically written into the brain and kept for decades is not understood.
Open in the interactive tree →Memory is thought to be stored in changes of synapse strength and in sparse groups of 'engram' cells. In 2012 the group of Susumu Tonegawa reactivated a fear memory in mice by switching on tagged neurons with light, and in 2013 false memories were implanted in mice. But how a memory persists for decades while molecules turn over, how it is recalled and how its content is represented remain open.
As of October 2026
Connectomes now show where synapses are, but not their strengths or history, and the MICrONS map covers just a cubic millimeter of mouse cortex. Engram experiments work in mice with simple memories; nobody can read out or write an arbitrary human memory. How a memory is physically formed and kept for decades is still an open question.
What is missing
- Methods to read out synapse strength and molecular state across whole circuits
- Long-term recordings of the same cells during learning and recall
- A theory of how distributed networks store, update and retrieve content
- Understanding of memory consolidation during sleep
- Human data: non-invasive tools lack resolution, and invasive access is rare
Becomes possible once solved
- Treatment of memory loss
- Memory prostheses
- Better learning methods and trauma therapy
Open steps
- Reading synapse strength Medium AI leverageRead out synapse strength and molecular state across whole circuits, not only where synapses sit.
- Inferring learning rules High AI leverageInfer how synapses change during learning from long recordings of the same cells during learning and recall.
- A theory of memory storage Medium AI leverageA theory of how distributed networks store, update and retrieve content over decades.
- Memory consolidation in sleep Medium AI leverageUnderstand what the brain replays during sleep and how it turns fresh memories into lasting ones.
- Human memory recordings Low AI leverageHuman data at cellular resolution: non-invasive tools lack resolution, and invasive access is rare.
Where AI could help
Medium AI leverage. AI automates wiring-map reconstruction, but synapse strength, molecular state and a theory of storage must come from new measurements.
- Segmenting neurons and detecting synapses in electron-microscopy volumes
- Inferring learning rules by fitting models to long recordings of the same cells
- Decoding what neural activity represents during learning and recall
- Building network models that test theories of memory storage and consolidation
Shown so far
- In April 2025 the MICrONS consortium published a cubic-millimeter mouse cortex map with about 200,000 cells and 523 million synapses, reconstructed with machine-learning segmentation plus human proofreading. source
- In December 2023 a Janelia-led preprint showed that gradient-descent fitting to recorded neural or behavioral data can infer synaptic plasticity rules, and found an active forgetting component in fly reward learning. source
Prerequisites
- Neuron doctrine1888
- Patient H.M. and memory1953H.M. showed which brain structures memory depends on
- Place and grid cells1971Place cells are the best-studied example of a memory-related code in neurons
- Long-term potentiation1973LTP is the leading cellular candidate for a memory trace
- Large-scale neural recording1990–2017
- Optogenetics2005
- Brain wiring maps2024