Stanford memory system recalls long histories with 16 memory tokens
A Stanford system compresses conversation and document history into 16 memory tokens and cuts document-token use by up to 4,212x.
According to an X post by Adib Vafa, Stanford researchers describe a memory system that recalls information from entire conversation histories using only 16 memory tokens. The headline figure is a reduction in document-token usage of up to 4,212x. In a 512-document test, the method ran 27x faster and outperformed existing approaches on 5 of 8 benchmarks.
Discussion of the work noted that compressing history this far is lossy. The open question is whether the right detail survives until it is needed.
Why it matters
If the efficiency holds, long-running agents could carry context at a fraction of the token cost, but whether they recall rarely used details is the real test.
What you can do
Track the paper and test recall on your own long-context tasks before changing memory architectures.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works