---
title: "Evaluate QA memory and action caching on your application"
description: "Compare controlled agent-qa runs with and without memory or cache while preserving fixture state, defect detection and comparable execution evidence."
canonical_url: "https://vostride.com/agent-library/software-factory/evaluate-qa-memory-and-cache.md"
md_url: "https://vostride.com/agent-library/software-factory/evaluate-qa-memory-and-cache.md"
content_kind: "guide"
topic: "software-factory"
---

# Evaluate QA memory and action caching on your application

Memory and action caching are useful only if repeated execution still checks the intended behavior. Evaluate them against your application's known outcomes, including at least one defect the test must reject. A faster passing run alone does not establish that the verification became more useful.

## Hold the experiment steady

Record the model, provider configuration, application revision, fixture snapshot, target and exact test definition. Use a disposable environment where the same initial state can be restored. Keep the required outcome fixed while varying the execution mode.

The [evaluation guide](/docs/agent-qa/guides/evaluating-agent-qa.md) documents relevant controls and interpretation. In the documented CLI, `--no-cache` and `--no-memory` disable their use for a run; those switches do not prove that the workspace started with an empty knowledge store. A cold-start experiment requires a fresh isolated workspace with its own runtime and memory paths.

## Compare correctness before speed

Run the unchanged scenario and a known failing scenario under each condition. Check that the failure is detected at the expected assertion. If one mode passes the defect, investigate the observed steps before counting any time or token savings.

Restore fixtures between runs and vary order to reduce warm-up bias. Record completed-run durations, reported model usage, retries and setup failures separately. Do not combine schema validation, blocked runs and actual application execution into one success percentage.

## Interpret the result narrowly

Report the number of runs and which application journeys were tested. Explain whether an improvement concerns setup, repeat planning or application interaction. Avoid extrapolating a result on one stable flow to a changing mobile application or a different model.

Keep the baseline and memory-enabled evidence so another engineer can inspect why the paths differed. Use the [memory maintenance reference](/docs/agent-qa/memory/reliability-maintenance.md) when investigating stored observations; avoid deleting knowledge merely to hide an unexplained failure.

[Caching reference](/docs/agent-qa/caching.md) · [Release evidence contract](/agent-library/software-factory/release-evidence-contract.md)