This dataset uses a search across compute-in-memory, in-memory computing, processing-in-memory, and analog in-memory compute terminology, restricted to IPC classes G11C7, G06F15 and G11C13. That means it captures memory-array circuit claims, digital data-processing claims, and static/dynamic memory claims that specifically implement or support in-memory computation. It excludes general memory patents that do not touch compute functionality, and it will miss filings that use different terminology and were classed outside those three IPC codes, so any single search string should be treated as one lens rather than the full picture.

A single invention can generate multiple documents through continuations, divisionals, and parallel filings in different jurisdictions, which inflates raw document counts without reflecting additional invention. Counting by patent family, as this landscape does with 2,460 families, neutralises that effect and gives a fairer sense of how many distinct inventions are actually being protected. It is particularly important in CIM, where large semiconductor assignees routinely file continuation chains around a single core architecture.

The trend shows growth from 69 families in 2017 to a peak of 389 in 2024, but the midpoint year 2022 already reached 337, and the two years around the peak show a flattening rather than continued acceleration. The apparent drop to 47 in 2026 is partly a data artefact: patent publication lags filing by roughly 18 months, so the most recent one to two years in any filing trend always look thinner than they eventually will. The honest read is a plateau since around 2022, pending confirmation once 2025–2026 data fully publishes.

Filing activity in the latest tracked year is led in absolute terms by Peking University at 5 filings, though that is down 64% year-on-year, alongside foundry and memory-vendor names each filing in the low single digits. Nearly every tracked assignee shows a year-on-year decline in the most recent partial year, with the notable exception of Qualcomm, which doubled its count to 2 filings. That pattern — high concentration at the top, broad cooling beneath it — suggests the field's most active filers are consolidating existing claim positions rather than opening new ones.

The core claim space around static/digital memory circuitry (G11C) and general digital processing (G06F) is dense, with 1,828 and 1,445 records respectively, so filing directly into that core is likely to run into crowded prior art. Thinner areas include reference-voltage generation circuitry for CIM arrays, global charge-sharing accumulation schemes, and coding/conversion approaches at the memory-compute boundary classed under H03M, which has only 67 records. These are not guarantees of patentability, but they are areas where the filing density is measurably lower than the core.