Percepta unveils Spotlight, an architecture that splits memory from intelligence

On October 3, 2026, a post on Reddit's LocalLLaMA community reported that Percepta has introduced a new architecture called Spotlight. According to the report, Percepta describes Spotlight as replacing attention with a memory that escapes the usual trade-off: it is said to be the first architecture to achieve infinitely growing memory without increasing the access cost.

The mechanism described is that every token reads from and writes to an unbounded memory, while the model learns to index individual memory units so that each token only touches a small number of units at a time. On this account, knowledge and skills can keep growing without any changes to the model weights.

The report points to two Percepta blog posts, titled "Can LLMs grow their own capabilities?" and "Spotlight memory." Both are attributed to Percepta's own site, and the quoted description of Spotlight as the first architecture to reach infinitely growing memory without raising access cost appears in that material.

That is where the story currently stands. The architecture was announced by Percepta, the source is Percepta's blog, and the coverage comes from the LocalLLaMA community on Reddit. No specific model name, version number, parameter count, benchmark result, or third-party verification appears in the material, and no follow-up beyond the announcement is described.