The article proposes moving memory storage inside the agent's reasoning loop to create an "extended working memory," arguing that latency is determined by the store's location rather than the retrieval pattern. By using an in-process store that answers in approximately 100 microseconds, the system avoids the high latency of networked stores which can inflate end-to-end time by up to 83x.
- Redundant actions drop from 7.2 out of 12 with a 110ms cloud round trip to zero at in-process speeds.
- Recall improves from 0/5 to 3.6-4.8/5 across four GPT-5-class models under a bounded window.
- The store maintained 100% write persistence (244 of 244 writes kept) with all misses attributed to the agent's read policy.
- Pairing the in-process store with a small local embedder reduces the dominant embedding cost, bringing the complete operation to approximately 40 microseconds.
This approach demonstrates that fast, constant availability transforms memory into extended working memory, significantly improving recall and efficiency compared to traditional external stores.