Research

Evaluating retrieval when the ground truth moves

Evaluating retrieval when the ground truth moves

Evaluating retrieval when the ground truth moves

Static benchmarks stop working when answers change. This is how we test retrieval against a moving target.

AR

Priya Nair

Jan 8, 2026 · 10 min read

·

Share

Most retrieval benchmarks assume a fixed set of documents and a fixed set of right answers. Company knowledge is neither. Prices change, owners change and last quarter’s decision gets reversed.

Questions with an expiry date

We label every evaluation question with the period in which its answer is valid. A retrieval system is scored on returning the answer that was true at the time of the question.

What we measure

Precision at five tells us whether the right memory is near the top. Staleness rate tells us how often a superseded memory outranks its replacement. Citation accuracy tells us whether the source link actually supports the answer.

Reading the results

The surprise was how often a high precision score hid a high staleness rate. A system can find a relevant memory and still find the wrong version of it.

AR

Written by Priya Nair

Head of Product

Follow on X

LinkedIn

Give your AI a memory.

Deploy in minutes, keep your stack, and stop repeating yourself. Free for the first 10,000 memories.

SOC 2 Type II · GDPR ready · Self-host available

Give your AI a memory.

Deploy in minutes, keep your stack, and stop repeating yourself. Free for the first 10,000 memories.

SOC 2 Type II · GDPR ready · Self-host available

RECALL

The cross-tool memory layer for AI. Built for teams who ship with agents, copilots and automation.

© 2026 Recall Labs, Inc.

Created by Svorla

RECALL

The cross-tool memory layer for AI. Built for teams who ship with agents, copilots and automation.

© 2026 Recall Labs, Inc.

Created by Svorla

Get Template

Create a free website with Framer, the website builder loved by startups, designers and agencies.