Just-in-Time Memory: saving raw task histories instead of end-of-run summaries
A post about Salesforce AI Research’s Just-in-Time Memory describes keeping raw task histories, then extracting relevant information when the next task arrives.
TLDR
A post describing Salesforce AI Research’s paper says Just-in-Time Memory uses a curator to read retrieved task histories alongside a new task and create a short memory tailored to it. Because that memory is used immediately, the curator can be trained on whether the same task succeeds. The post reports gains over the strongest baseline of 16.2 success-rate points on ALFWorld, 16.3 on WebShop and 3.9 on tau2-bench. It also says the untrained curator matches or beats memory written when a task ends.
