AI Agent News/stories/stanford-researchers-release-memorilla-a-latent-memory-module-for-frozen-llms.md

Memory

Stanford researchers release Memorilla, a latent memory module for frozen LLMs

Stanford researchers have released Memorilla, an open-source memory system that compresses a document collection into 16 memory tokens for a frozen Qwen3-8B model. The source reports that it reaches 84.33% on the FactKG benchmark while using 108 times fewer document tokens than full-context input.

Source: www.marktechpost.com/2026/11/14/stanford-memorilla-memory-system-16-memory-tokens

Memorilla is a memory system that the Stanford researchers have made available as open source. It works by compressing a document collection into 16 memory tokens, and the underlying language model stays frozen, meaning its weights are not changed.

The system is applied to Qwen3-8B. According to the report, it reaches 84.33% on FactKG while using 108 times fewer document tokens than a full-context approach.

The report is short and gives no further details on training, latency or how the memory tokens are produced, so those points would need to be checked in the released code and its documentation.

Why it matters if you build agents

Developers who run agents over large document sets could test whether a compact memory representation can replace sending full documents into the context window on every call.

Because the reported gains come from one model and one benchmark, the results should be checked on your own data before you change an agent's retrieval or memory design.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe