Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

memory-hygiene-bench — measuring the cost of dirty agent memory

memory-hygiene-bench is an open benchmark that measures the cost of degraded AI agent memory and the return from consolidation: agent task success, token cost, and repeat-failure rate across three arms — no memory, degraded memory, and consolidated memory.

AI agent memory degrades over time — duplicates, contradictions, stale facts, and poisoned entries accumulate — and existing benchmarks do not price that degradation. This benchmark is an intervention axis, not another memory benchmark: existing benchmarks measure whether memory helps (STATE-Bench, MemoryArena, AMB) or whether agents resist stale memory (MemSyco-Bench); this one measures what memory degradation costs and what memory hygiene returns. It is designed to run on top of existing open environments (STATE-Bench, MemoryArena) rather than replacing them.

Three-arm design

Every task suite runs three times with identical tasks and seeds:

  • Arm A — no memory: the agent starts cold; baseline for what memory is worth at all.
  • Arm B — degraded memory: the agent runs with accumulated dirty memory (duplicates, contradictions, stale and injected entries, seeded from a published error taxonomy with ground truth).
  • Arm C — consolidated memory: the same memory after cleanup and consolidation.

Metrics per arm: task success rate, tokens per task, and repeat-failure rate (does the agent make the same mistake twice).

Honesty rules

The benchmark's maintainers build a consolidation tool, so the rules are strict by construction:

  • Open code, open data, fixed seeds.
  • One-command replication of every reported number.
  • External memory and consolidation systems are welcome and can be plugged into arms B/C.

Related benchmarks

Status

Early development. This repository reserves the benchmark's public home; the task suites and harness follow.

Project family

License

Dual: code under Apache-2.0, data under CC BY 4.0.

About

An open benchmark measuring the cost of dirty agent memory.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors