A quantum memory can answer your query without learning what you asked

Ask a remote server to look something up for you, and the server learns what you were curious about. That leak is annoying with a search engine and dangerous with a medical database. It is also a problem quantum computers inherit rather than escape, because many quantum algorithms lean on a subroutine that does exactly this: reach into a memory and pull out an entry, in superposition, across many addresses at once.

The hardware for that job is called Quantum Random Access Memory, or QRAM, and it lets a program query a stored database coherently, meaning the query can address many locations simultaneously and keep the quantum interference that makes the algorithm fast. Building one is hard enough that it is easy to imagine a future where a client with a modest quantum device rents time on a much larger machine elsewhere. Shifan Xu, Yizhuo Tan, Yongshan Ding, and Jakub Szefer, writing in a preprint posted to arXiv on 29 July 2026, point out the catch. Hand your query to someone else's QRAM and the standard interface hands them your secrets along with it.

Their answer is a construction they call oblivious QRAM, or OQRAM. The name borrows from a well-established classical idea, oblivious RAM, where a client shuffles and encrypts stored data so that the pattern of memory accesses reveals nothing about what is being accessed. The team's contribution is a version of that idea that survives contact with quantum queries, where the address itself is not a single number but a superposition that must stay intact for the algorithm to work.

How the hiding works

The protocol splits into two parts. An offline refresh phase stores the database on the server in encrypted and shuffled form, so the physical layout carries no information about the logical one. Then, during the online phase, the client masks the address of each query before sending it.

Masking a quantum address is the delicate step, since scrambling it must not destroy the superposition. The authors offer two routes. One uses a quantum-secure pseudorandom permutation, a scrambling function that stays unpredictable even to an adversary with a quantum computer. The other uses a quantum one-time pad, the quantum analogue of the classic unbreakable cipher in which a fresh random key is used exactly once. Both are applied coherently, so the mask travels with the superposition rather than collapsing it.

What makes the accounting interesting is the cost. A client who wanted this privacy guarantee without a server could simply build a local QRAM, but the authors note that the quantum resources for an equivalent local construction grow exponentially with the size of the address, which is another way of saying the whole point of delegating is that you cannot afford the hardware. Under the client model the team adopts, the online protection adds only what they describe as modest quantum overhead beyond the query register itself. Compared with fully blind quantum computing, the heavier and more general framework in which a client hides an entire computation from a server, they report an exponential reduction in quantum communication, and lower quantum resource requirements on both ends.

The qPRP variant has a second practical feature. Refreshing a shuffled database is expensive in classical communication, so the protocol spreads that refresh across many queries rather than paying it all at once, which makes repeated use more affordable.

A server that merely watches is one threat. A server that lies is another, and the authors address it with decoy checks: the client mixes in queries whose correct answers it already knows, then verifies them. Tampering is caught probabilistically, meaning a dishonest server might get away with a single alteration but is increasingly likely to be caught as the checks accumulate. The paper frames this as strengthening privacy against malicious servers rather than eliminating the risk.