Magneto-optical memory stores bits in the magnetisation of a material and writes or reads them with light on a photonic chip; cell-level demonstrations and wafer-scale garnet films on silicon now exist, but nothing resembling an addressable array does.
Summary
Magneto-optical (MO) memory combines two mature ideas. The storage element is magnetic and therefore non-volatile: a magnetic tunnel junction, a ferrimagnetic garnet film, or a two-dimensional magnet. The access mechanism is optical: light either flips the magnetisation directly (all-optical switching, typically with femtosecond pulses) or senses it through the Faraday and Kerr effects, in which magnetisation rotates the polarisation of transmitted or reflected light. Put on a photonic integrated circuit, the appeal is that a photonic processor could keep its weights or state in place without converting to the electrical domain to read or write them.
Viability (2/5)
TLDR: Working single cells and wafer-scale films; no addressable array and no solution to sub-diffraction addressing.
Drivers (2/5)
On the supply side there is a coherent push: wafer-scale garnet on silicon with the metrology needed to qualify it, a steady flow of magnetic-materials work that improves the underlying switching physics (strain tuning of demagnetisation in a room-temperature van der Waals ferromagnet, with 1.2% tensile strain cutting demagnetisation time by about 20%; magnetically switchable chiral second-harmonic emission with remanent states in bilayer CrSBr), and adjacent chip-scale light modulation capability such as a non-suspended push-pull TFLN acousto-optic modulator at 1.004 V cm and 132.5 MHz bandwidth.
On the demand side the evidence points elsewhere. Money is flowing into photonic compute boxes with integrated memory (Olix raised $220M Series A for an optical tensor processing unit with integrated memory ref) and into electrically-addressed MRAM, where Everspin reported 238 design wins in 2025 ramping to production ref. Neither is a customer asking for optically addressed magnetic memory; both are plausible acquirers of the IP. Score 2 because pull is inferred, not observed.
TLDR: Supply-side materials and metrology progress is genuine; the sources show no demand pull for MO memory specifically.
Novelty (3/5)
TLDR: One quantified 100x speed advantage over incumbent photonic memory, with an unnamed comparator and no comparison against electronic memory.
Diffusion (2/5)
TLDR: The garnet process is heading for isolators, not memory; MO memory has no product channel in the sources.
Impact (unscored)
A score here would be invented. It is left null deliberately. The specific numbers that would allow one are listed in the open questions.
TLDR: The sources do not size the system-level gain or the market, so no defensible score.
Timing Later (5-10yr)
Even on an optimistic reading in which that milestone clears in 2027, the path from a sub-diffraction single-cell write to an array with characterised read margin, retention and yield is a multi-year programme, and the sources give no evidence that any organisation is running one. Later (5-10yr) is the honest band for a memory product; anyone promising sooner should be asked for the addressing scheme.
Overrated or underrated? Too early to say
Prediction
Evidence base
Open questions
Assessment drafted 2026-08-31 from up to 11 KB sources using the technology-scorecard framework; scores are a draft read pending review.