StarlightSearch Launches Reflect: Utility-Ranked Memory System for Self-Improving AI Agents

New approach closes the feedback loop between agent observability and performance, enabling continuous improvement without prompt engineering

StarlightSearch, a startup building infrastructure for self-improving AI agents, today announced the launch of Reflect, a utility-ranked memory layer that ranks retrieved guidance by actual outcomes rather than semantic similarity alone.

The announcement addresses a persistent gap in production AI systems: while most organizations now have robust observability stacks capturing agent traces and evaluation frameworks measuring pass/fail rates, these systems rarely connect. Agents start each task from a blank slate, unable to learn from previous failures.

“Every AI team we talk to has the same frustration,” said Sonam Pankaj, founder of StarlightSearch. “They can see exactly where their agents fail. They have dashboards full of traces. But turning those failures into better behavior requires manual intervention. We built Reflect to automate that learning loop.”

Also Read: AiThority Interview with Glenn Jocher, Founder & CEO, Ultralytics

How Reflect Works: The Utility Difference

Traditional memory systems for large language models rely on semantic similarity: they retrieve content that sounds relevant to the current query. Reflect adds a second dimension — utility, a score that tracks whether following a particular piece of retrieved advice actually led to success.

The system uses a weighted scoring formula where the score balances semantic relevance against proven effectiveness. A memory that has been retrieved multiple times and consistently contributed to successful outcomes will rank higher than one that merely sounds similar to the current task.

Think of it like a credit score,” Pankaj explained. “It doesn’t just record that you had a loan. It tracks whether you paid it back. Similarly, utility tracks whether a memory actually helped the agent succeed.”

From Facts to Reasoning

Unlike conventional memory layers that store static facts — user preferences, document chunks, conversation history — Reflect stores reasoning about outcomes. When a trace is reviewed (marked as pass or fail), Reflect generates a reflection: a compressed lesson about what went wrong and what should be done differently next time.

For example, a customer support agent handling refund requests might initially retrieve advice to “issue an immediate refund” for duplicate charge complaints. If that leads to a double-refund because settlement status wasn’t checked, the utility of that memory drops. Simultaneously, Reflect stores a new reflection: “For duplicate charge complaints, check settlement status before initiating any refund.” On subsequent similar tickets, the new reflection ranks higher while the old one is deprioritized — all without human prompt engineering.

Three-Layer Integration

Reflect sits between three existing components in production agent architectures:

– Observability: Traces capture every tool call, LLM completion, and exception
– Evaluation: Reviews mark outcomes as pass, fail, or provide detailed feedback
– Action: The agent retrieves memories before executing

The company’s approach treats traces not as passive audit logs but as training signal. When a review marks a trace as failed, the system extracts a reflection and stores it as task-linked memory. When a similar task arrives, that reflection surfaces with an updated utility score.

“What makes this production-ready is that it’s outcome-addressable,” Pankaj said. “You’re not retrieving by keyword. You’re retrieving by semantic similarity weighted by whether those memories have historically helped or hurt. The eval outcome becomes a first-class signal in retrieval ranking.”

Market Context

The launch comes as enterprises increasingly deploy AI agents for customer support, code review, and task automation — use cases where consistent, reliable behavior matters more than one-off heroic performance.

Existing memory frameworks have focused primarily on user continuity: personalization, preferences, and conversation history. Academic work including Reflexion demonstrated that agents can learn from verbal self-reflection, achieving 91 percent pass rates on coding benchmarks. However, these approaches do not incorporate learned utility signals that rank which experiences to surface.

Reflect’s differentiation lies in its quantitative ranking layer. While semantic memory retrieves what sounds relevant, Reflect retrieves what has earned trust through repeated reviewed runs.

Also Read: ​​The Infrastructure War Behind the AI Boom

[To share your insights with us, please write to psen@itechseries.com ]