Skip to content

Ablation Tests

Every core module in Mnemo has been validated through ablation experiments — systematically removing each component and measuring the impact on retrieval quality.

35 tests, 3 suites, 100% pass.

Why Ablation?

Most memory frameworks claim features without proving they help. Mnemo takes a different approach: every module must earn its place through measurable contribution. If removing a component doesn't degrade performance, it shouldn't exist.

Results Summary

SuiteTestsKey Finding
Decay & Tier system12Each tier provides measurable score separation; flat models lose discrimination
Emotional salience12Removing salience collapses ranking spread to zero — essential signal
Parameter A/B test11Current parameters confirmed optimal via 4-horizon comparison

Key Findings

  1. Every module is load-bearing — no component can be safely removed without measurable quality loss
  2. Parameters are at the sweet spot — A/B testing confirmed no adjustment needed
  3. Tier system is essential — flat decay loses both preservation of important memories and cleanup of noise
  4. Emotional salience is not optional — without it, the system cannot distinguish emotional from routine memories

Running the Tests

bash
node --test packages/core/test/reflection-ablation.test.mjs
node --test packages/core/test/salience-ablation.test.mjs
node --test packages/core/test/salience-ab-test.mjs

Released under the MIT License.