Where are old memories protected when we learn something new?

Synapses change when we learn a new experience. If that change extends to synapses that carry an earlier experience, new learning could interfere with an old memory.

Onasch and colleagues asked whether shared neurons might avoid interference by using different dendritic branches for learning. They investigated this with a computational model that includes dendrites and inhibitory circuitry. Neuron study ↗

Each model neuron has several dendritic compartments. Synaptic strength changes in response to local activity on each branch. Changes are less likely on strongly inhibited branches.

The model relieves inhibition on selected branches according to context. Connections needed for new learning can change, while connections previously formed on other branches are protected.

Learning overlapping inputs in succession can weaken connections formed earlier. Switching the branches available for learning when the context changed reduced that interference. A neuron can take part in several memories while the synaptic changes occur in different parts of the cell. Authors’ analysis notes ↗

This raises a question about the scale at which a memory is protected.

At the level of whole neurons, the same cells may be active for old and new experiences. Yet the branches that receive input and the connections that change with learning can differ. Overlap among active cells can coexist with protection of the connections that support a memory.

In our article on representational drift ↗, we asked whether information could still be read as neural activity changes. The new idea is that change may be uneven, with some parts of the circuit protected from modification.

Dropout raises a related question about how a system maintains its output when some elements are unavailable. Here, the question is which connections are available for modification during new learning. Filling in missing information and avoiding interference during learning could be separate contributors to memory stability.

One test in a living brain would be to disrupt branch-selective inhibition only while a new task is learned. If the older memory remains inaccessible after inhibition is restored, the disruption may have removed protection during learning. If recall returns, the effect may instead have been on access to the stored memory.

The context signal in the model comes from outside the network. How a brain distinguishes experiences and assigns branches to learning remains open. The model gives us a specific hypothesis to investigate alongside the question of which cells carry a memory: how the circuit chooses where to allow change.