A utility-ranked memory method proposes using a history of successes and failures to improve runtime performance.
Most AI agent systems in production operate with a structural deficiency: each new execution starts from scratch, failing to leverage learning from previous attempts. According to Sonam Pankaj, CEO and co-founder of StarlightSearch, while observability tools record execution traces and evaluation systems log successes and failures, the agent running on a given day retains no memory of why the previous day's executions succeeded or failed.
Current memory approaches have specific limitations that prevent this learning loop from closing. Conversation buffers consider only information recency; semantic systems retrieve content based on textual similarity rather than proven utility; and reflection-based methods capture lessons without distinguishing which ones actually work in practice. The gap between observation and action remains open.
Pankaj's proposal introduces the concept of utility-ranked memory, which treats each memory like a credit score. When a memory is retrieved and the agent's execution succeeds, its utility increases; when the execution fails, its utility decreases. The ranking formula combines semantic similarity with the history of outcomes.
According to the presentation, the method was demonstrated using a SQL agent that updates its context at runtime based on the outcome of each operation. The entire update process occurs during execution, without manual intervention. Pankaj is also the co-creator of embedanything, a Rust-based pipeline for RAG (Retrieval-Augmented Generation) with over 450,000 downloads and contributions from companies such as Elastic, Milvus, and Qdrant.
The core issue raised is that utility should be the primary criterion for agent memory, rather than mere similarity or recency. The argument is that without a mechanism to weigh memories by their history of outcomes, agents repeat mistakes and fail to consolidate successes, limiting performance evolution over time.
Most AI agents start each new execution from scratch, lacking a mechanism to retain and learn from the successes and failures of previous runs. This causes them to repeat mistakes and fail to consolidate successes over time.
Utility-ranked memory is a method that treats each memory like a credit score. When a retrieved memory leads to a successful agent execution, its utility increases; if the execution fails, its utility decreases. It combines semantic similarity with the history of outcomes.
Current approaches rely on conversation buffers (which only consider recency), semantic systems (which retrieve based on textual similarity rather than utility), and reflection-based methods (which capture lessons without distinguishing what actually works in practice).