EverMind released a research paper describing a 100-million-token long-term memory framework using Memory Sparse Attention architecture [1].

This development addresses a primary limitation in large language models by allowing them to process and recall vastly larger amounts of information. By breaking previous token limits, the technology enables AI to maintain context over much longer sequences of data without the typical performance degradation associated with massive context windows.

The company, based in San Mateo, California, said it announced the architecture in a press release on March 18, 2026 [2]. The Memory Sparse Attention (MSA) framework is designed to provide efficient end-to-end long-term memory for LLMs [2].

According to technical documentation, the MSA architecture successfully achieved a token context length of 100 million tokens [3]. This capacity allows a model to ingest and reference an expansive dataset, such as entire libraries of technical manuals or massive codebases, within a single operational window.

While the industry has seen various attempts to increase context windows, EverMind's approach focuses on the efficiency of sparse attention. This method reduces the computational overhead required to track dependencies across millions of tokens, which has historically been a barrier to scaling AI memory.

The release comes amid a broader trend of open-weights competition in the AI sector. For example, the company MiniMax said it committed to an open-weights release on June 1 [4]. These movements suggest a shift toward more transparent and accessible high-capacity architectures in the U.S. tech corridor.

“a 100-million-token long-term memory framework”

The move toward 100-million-token windows represents a transition from 'short-term' prompt engineering to true long-term digital memory. If this architecture scales efficiently, AI will no longer require frequent summaries or external databases to remember a user's history, potentially enabling models to act as comprehensive personal or corporate knowledge bases with perfect recall.