Why is the cortex a slow learner?

Article metrics

Mice that have only half the normal amount of a synaptic protein called α-CaMKII learn normally, but remember poorly. The result sheds light on the mysterious mechanisms of memory consolidation.

The gene-knockout business — engineering mutant animals with specific gene deletions — has turned up many oddities. Most are destined to be tangents on the path to discovery, but a skilful scientist can sometimes capitalize on unexpected findings. So it was with Frankland and colleagues' work on the involvement of an enzyme called α-calcium/calmodulin-dependent protein kinase type II (α-CaMKII) in neuronal plasticity and learning. As they describe on page 309 of this issue1, they were studying α-CaMKII heterozygous mice, in which just one of the two copies of the gene encoding α-CaMKII was knocked out. Focusing on long-term potentiation (LTP), a molecular mechanism for strengthening neuronal connections that is widely held to be the basis of memory formation2, the authors observed that this process was induced normally in both the hippocampus and the cortex in the brains of the heterozygous mice. But LTP then decayed unexpectedly quickly in the cortex, while persisting normally in the hippocampus.

These findings were brought to bear on the 'consolidation' theory of memory3,4. According to one mechanistic explanation5 of this idea, newly acquired sensory information is funnelled through the cortex to the hippocampus. Surprisingly, only the hippocampus actually learns at this time — it is said to be online. Later, when the hippocampus is offline (probably during sleep), it replays stored information, transmitting it to the cortex. The cortex is considered to be a slow learner, capable of lasting memory storage only as a result of this repeated replaying of information by the hippocampus. In some views, the hippocampus is only a temporary memory store — once memory traces become stabilized in the cortex, memories can be accessed even if the hippocampus is removed6. There is now direct evidence that some form of hippocampal replay occurs7,8.

LTP is thought to be a molecular mechanism underlying learning, and requires the protein α-CaMKII (refs 910). This protein is greatly enriched at many excitatory neuronal connections (synapses) in the central nervous system. However, α-CaMKII heterozygous mice have only half the normal amount of α-CaMKII. Frankland et al.'s finding1 that these mice show normal LTP in the hippocampus, but transient LTP in the cortex, allowed them to study consolidation in a new way. According to the theory above, the heterozygous mice should learn normally at first. But because LTP in the cortex is transient, information should not be stabilized there. The animals would be expected to forget quickly, and this is exactly what Frankland et al. observed. In two 'hippocampus-dependent' learning tasks, the mice learnt normally at first, but then forgot things more quickly than normal mice.

An elegant feature of the study1 was the care taken, through further behavioural analyses, to explore alternative explanations of this apparent dissociation between initial learning and memory retention. One possibility was a 'ceiling' effect: the initial memory trace in the hippocampus might have been weaker in the heterozygous mice. However, when the authors trained the mice more intensively, the heterozygous animals still forgot more quickly. The implication is that trace strength and memory persistence are dissociable features of memory. This makes sense if they are, in part, stored in different parts of the brain.

These results support the idea that the hippocampus is the fast online learner that 'teaches' the slower cortex offline. But a close look at Fig. 3 on page 311 reveals an apparent contradiction: both the cortex and the hippocampus show rapid induction of LTP in response to brief stimulation, with LTP in the cortex remaining stable in normal mice. Why, then, should the cortex be a slow learner?

A possible solution relates to the fact that synaptic connections must already exist if they are to be strengthened by LTP. The most common form of LTP depends on NMDA-type receptors for the neurotransmitter glutamate to strengthen existing connections between neurons that fire simultaneously. As Hebb originally argued11, such modifications could account for those types of learning in which associations are made between events or items that occur at the same time. In neuronal networks in the hippocampus, it is reasonably likely that there is a pre-existing synapse between any two cells, because the number of contacts a neuronal axon can make (10,000) is only slightly smaller than the number of cells in the network (generally less than a million)12. So a sensory input would have a good chance of strengthening many synapses in the hippocampus by LTP. Networks in the cortex are much larger (within the 8-mm extent of horizontal cortical axons13, there are 10 million cells), and the probability of any two cells being connected is much lower. So there may be too few pre-existing connections that can be strengthened by LTP during learning.

But a connection between two cortical neurons might be formed and strengthened offline (during hippocampal replay) in a two-step process. The first step would be an ongoing process involving the perhaps random growth of new synapses and the breaking of older, non-stabilized ones. As the replay occurs repeatedly, when two cells do by chance become physically connected, they would then be sensitive to LTP if they were simultaneously excited as a result of hippocampal replay.

The observed development of new synapses in adult brains is consistent with this proposal. When animals are raised in mentally stimulating environments, the number of synapses increases14; studies of the mollusc Aplysia show a similar pattern15. More direct evidence comes from studies of horizontal axonal connections in the visual cortex of adult cats13 and in the monkey somatosensory cortex16. These studies show that changes in the input activity to the cortex can lead to increases in axonal branching and in the number of synaptic boutons (terminals). This suggests that new connections can be made in adult brains, and that these connections are determined by processes that depend on neuronal activity.

With the α-CaMKII heterozygous mice, it is now possible to manipulate cortical and hippocampal plasticity separately. It may also be feasible to settle some long-standing issues about how the processes in memory consolidation relate to retrograde amnesia. In humans, retrograde amnesia reflects the loss of access to, or the failure to consolidate, memory traces formed before brain damage. Animal models reveal a relationship between remembered events and the time before brain damage when those events occurred, providing hints to the underlying memory-consolidation process17. But such temporal gradients are not always seen, and certain features of retrograde amnesia can be understood instead in terms of the opportunities that recall allows for making several traces of one event, rather than the consolidation of a single trace18.

Timing the heterozygous deletion of α-CaMKII in mice would shed fresh light on this controversy. The mutant animals might remember information acquired shortly before the timed gene knockout. This would cast doubt on a slow, continuous interaction between brain areas and suggest that it is important that several memory traces are made in the hippocampus (in which LTP is normal in these mice). Alternatively, the mice might rapidly forget both old and new information. In either case, control of the mutation should enable us to probe deeper into the mysteries of memory.


  1. 1

    Frankland, P. W., O' Brien, C., Ohno, M., Kirkwood, A. & Silva, A. J. Nature 411, 309–313 (2001).

  2. 2

    Martin, S. J. et al. Annu. Rev. Neurosci. 23, 649–711 (2000).

  3. 3

    McGaugh, J. L. Science 153, 1351–1358 (1966).

  4. 4

    Squire, L. R. Psychol. Rev. 99, 195–231 (1992).

  5. 5

    Buzsáki, G. Neuroscience 31, 551–570 (1989).

  6. 6

    Eichenbaum, H. B. & Nadel, L. (eds) Hippocampus 11, 1–60 (2001).

  7. 7

    Skaggs, W. & McNaughton B. Science 274, 1870–1873 (1996).

  8. 8

    Louie, K. & Wilson, M. Neuron 29, 145–156 (2001).

  9. 9

    Kennedy, M. B. Nature 335, 770–772 (1988).

  10. 10

    Lisman, J. Trends Neurosci. 17, 406–412 (1994).

  11. 11

    Hebb, D. O. The Organization of Behaviour (Wiley, New York, 1949).

  12. 12

    Rolls, E. T. & Treves, A. Neural Networks and Brain Function (Oxford Univ. Press, 1998).

  13. 13

    Darian-Smith, C. & Gilbert, C. Nature 368, 737–740 (1994).

  14. 14

    Klintsova, A. & Greenough, W. Curr. Opin. Neurobiol. 9, 203–208 (1999).

  15. 15

    Bailey, C. et al. Nature Rev. Neurosci. 1, 11–20 (2000).

  16. 16

    Florence, S. L. et al. Science 282, 1062–1063 (1998).

  17. 17

    Anagnosteras, S. G. et al. J. Neurosci. 19, 1106–1114 (1999).

  18. 18

    Nadel, L. & Moscovitch, M. Curr. Opin. Neurobiol. 7, 217–227 (1997).

Download references

Author information

Rights and permissions

Reprints and Permissions

About this article

Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.