Amyloid has a bad reputation. In neuroscience and in public understanding, amyloid proteins are associated with pathology: the plaques that accumulate in Alzheimer’s disease, the misfolded protein aggregates that disrupt synaptic function and kill neurons. The implicit assumption has been that any amyloid formation in the brain is something going wrong, a sign of disease rather than a feature of normal biology.

Research published by the Stowers Institute for Medical Research in January 2026 overturns this assumption in a specific and important way. The work provides the first direct evidence that the nervous system deliberately forms amyloid-like protein structures as a mechanism for storing long-lasting memories. The proteins in question are not pathological aggregates accumulating by mistake. They are intentionally synthesized and maintained by neurons as a physical substrate for persistent memory traces.

The finding has implications well beyond neuroscience. For anyone working on whole brain emulation, it raises a question that had not been seriously posed before: if memories are partly encoded in amyloid protein structures, and not only in synaptic weight matrices, what does that mean for what a brain scan must capture to preserve a mind?

The Discovery

The Stowers Institute team was studying memory consolidation in animal models, examining what molecular changes accompany the transition from short-term to long-term memory storage. The distinction between the two is well established at the systems level: short-term memories are labile, easily disrupted, and do not require new protein synthesis. Long-term memories require gene expression and new protein production — a process called consolidation — and are resistant to disruption once consolidated.

The molecular identity of the proteins involved in long-term storage has been an open question for decades. Synaptic potentiation (the strengthening of connections between neurons) is one part of the answer, but it has been understood for some time that synaptic weight changes alone do not fully account for the persistence of memory across biological timescales. Synaptic proteins turn over continuously — individual synaptic proteins have half-lives of hours to days. Yet memories can persist for decades. Something must maintain the memory trace through this protein turnover.

The Stowers team identified a protein that forms amyloid-like aggregates specifically in neurons that have undergone learning-associated activity. The aggregates are found at synapses involved in the learned behavior. They are not present in naive tissue or in neurons that did not participate in the learning event. When the aggregates are experimentally disrupted, long-term memory is impaired. When aggregation is promoted, memory persistence is enhanced.

Critically, the aggregates show the hallmark biochemical features of amyloids — cross-beta sheet structure, resistance to denaturation, self-templating propagation — but are composed of a functional neural protein, not of misfolded disease-associated proteins like amyloid-beta or tau. The distinction is the same as the distinction between prion proteins in healthy brain (where they serve normal functions) and disease-causing prions (where misfolding leads to pathology). The architecture is similar; the identity and regulation of the protein determines whether it is functional or destructive.

Why Amyloids Are Suited for Long-Term Storage

The biophysical properties of amyloid structures explain why they are, in retrospect, a plausible substrate for persistent memory. Amyloid fibers are among the most stable protein structures known. They resist denaturation by heat, detergent, and proteolytic enzymes. Once formed, they serve as templates for incorporating additional monomers, allowing them to grow and persist against the background of protein turnover.

These properties are directly relevant to the memory persistence problem. If a synapse needs to maintain a modified state for years while its component proteins are individually replaced every few days, the information must be encoded in a structure that is stable against that turnover. A self-templating amyloid fiber that recruits new protein monomers while maintaining its structural template is one solution. Each new monomer that joins the fiber adopts the fiber’s conformation, perpetuating the structural information regardless of which specific protein molecules are present at any given time.

The analogy to digital storage is imprecise but useful: the amyloid structure encodes information in its spatial organization (which monomer occupies which position in the fiber, with what neighbors) rather than in the identity of the individual molecular components. This is structurally analogous to how magnetic domains in a hard drive encode information in orientation rather than in atomic composition.

Implications for What Mind Uploading Must Capture

The conventional model of memory storage in the context of brain emulation has been primarily connectomic: memories are stored as patterns of synaptic weights, and capturing the connectome — the full graph of neural connections with associated synapse strengths — captures the memories. This model underlies most thinking about what a brain scan would need to achieve to support mind uploading.

The Stowers findings complicate this model in a specific way. If memories are partly encoded in amyloid protein structures at synapses, then capturing synaptic connectivity alone may be insufficient. The amyloid substrate adds a molecular layer of information on top of the connectivity layer. Two synapses with identical connectivity geometry and receptor composition might differ in their amyloid state — one has an established amyloid scaffold encoding a learned association, the other does not.

Whether this additional layer of information is necessary to reconstruct a functionally accurate memory depends on what the amyloid scaffold encodes. If it primarily regulates synaptic strength (acting as a persistent molecular signal that maintains potentiation), then its functional contribution might be adequately approximated by measuring the synaptic weight at the time of scanning. If it encodes structural information not represented by synaptic weight — the pattern of plasticity, the sequence in which a learning event occurred — then it constitutes a separate channel of information that would need to be captured independently.

This distinction matters enormously for the bandwidth bottleneck analysis of memory and BCI: even if a future scanning technology can read synaptic weights with high fidelity, it may need an additional molecular imaging capability to read amyloid states at individual synapses. Current electron microscopy can detect amyloid fibers morphologically, but the protein identity of the fibers requires molecular contrast methods (immunolabeling or mass spectrometry) that are not compatible with standard connectomics workflows.

Relationship to Alzheimer’s Disease Research

The discovery that functional amyloids exist alongside pathological ones raises a practical challenge for Alzheimer’s research. Most amyloid-targeting therapeutic strategies have been based on the assumption that reducing amyloid burden in the brain is safe and beneficial. Two recent high-profile Alzheimer’s treatments — lecanemab and donanemab — work by clearing amyloid-beta plaques.

If functional amyloids in neurons are necessary for memory storage, aggressive amyloid clearance therapies could, in principle, disrupt normal memory mechanisms alongside the pathological aggregates they target. The proteins involved are different — the functional memory amyloids identified by the Stowers team are not amyloid-beta — but the finding motivates caution and more specific targeting in therapeutic development.

For the memory storage research field, it opens a new set of questions: how many proteins form functional amyloids during learning? Are there specific neurons, circuits, or memory types more reliant on amyloid-based storage than others? Can amyloid state be pharmacologically modulated to enhance or suppress specific memories?

Implications for Brain Preservation

Brain preservation research is increasingly focused on preserving ultrastructure at the molecular level, not just the gross anatomy of connections. The Song et al. work demonstrated that pig brain ultrastructure can be preserved with near-perfect fidelity within a 14-minute postmortem window. The relevant question is now: what constitutes “sufficient” molecular preservation for the purposes of reconstructing cognitive function?

Amyloid structures are, as noted, among the most stable protein architectures. They are likely to survive standard fixation and preservation protocols used in connectomics, including glutaraldehyde-based chemical fixation and cryogenic preservation. This is a tentative reassurance — the amyloid scaffold information may be preserved in the tissue even without specific steps to protect it. But validating this against known amyloid-based memory markers in preserved versus fresh tissue is now a research priority that the Stowers findings have newly motivated.

Cryonics preservation approaches have long argued that ultrastructural preservation is sufficient to preserve the information content of a brain. The amyloid finding supports that claim for the specific case of functional memory amyloids, which are likely more stable than most other molecular memory candidates. But it also highlights that “ultrastructural preservation” must be interpreted broadly enough to include molecular-level protein conformation, not just membrane topology and synaptic geometry.

The Molecular Complexity Problem

The Stowers findings are one data point in a growing body of evidence that memory storage is molecularly more complex than the synaptic weight model suggests. Other candidates for molecular memory storage mechanisms include persistent post-translational modifications of synaptic proteins, RNA localization at synapses, and epigenetic modifications in neuronal nuclei.

None of these individually constitute a complete alternative theory of memory. The weight of evidence still supports synaptic strength changes as the primary mechanism for most forms of learning-dependent plasticity. But the accumulating evidence that multiple molecular layers contribute suggests that a complete model of memory will involve more variables than the connectome alone.

For whole brain emulation, this creates an expanding list of what must be measured. Synaptic connectivity and weights. Amyloid states at individual synapses. Post-translational modification patterns. RNA localization. None of these measurements are individually prohibitive given sufficient scanning technology. But together they raise the resolution requirements for what constitutes a complete scan of a brain’s information content.

TRL Assessment: TRL 2–3. The core finding of functional neural amyloids is established in animal models. The implication for memory encoding specificity (what information the amyloid structure encodes beyond synaptic weight) is not yet resolved. Human relevance has not been directly demonstrated.

Official Sources

• Stowers Institute for Medical Research (January 2026) — Functional amyloids as a substrate for long-term synaptic memory. MedicalXpress coverage: https://medicalxpress.com/news/2026-01-reveals-brain-memory-tiny-protein.html
• Crick, F.H.C. (1984) — Memory and molecular turnover. Nature, 312(5990):101. DOI: 10.1038/312101a0
• Si, K. et al. (2003) — A neuronal isoform of CPEB regulates local protein synthesis and stabilizes synapse-specific long-term facilitation in Aplysia. Cell, 115(7):879–891. DOI: 10.1016/s0092-8674(03)01020-1
• Bhanu, M.K. et al. (2021) — Phase separation and pathological aggregation in the nervous system. Nature Reviews Neuroscience, 22(8):447–468. DOI: 10.1038/s41583-021-00462-6
• Tonegawa, S. et al. (2015) — Memory engram storage and retrieval. Current Opinion in Neurobiology, 35:101–109. DOI: 10.1016/j.conb.2015.07.009
• Related: WARP: Whole-Brain Gene Expression and Neuronal Activity Co-Mapped in Zebrafish