---
pretty_name: FalseMemBench
license: mit
task_categories:
  - text-retrieval
language:
  - en
tags:
  - retrieval
  - memory
  - llm-agents
  - adversarial
size_categories:
  - n<1K
---

# FalseMemBench

This dataset contains adversarial distractor cases for evaluating memory retrieval systems used by LLM agents.

Each case contains:

- a query
- a small candidate corpus
- one or more relevant entry ids
- an adversary type label

The dataset is intended for evaluating retrieval under confusion pressure rather than open-ended generation.

Current adversary types:

- entity swap
- environment swap
- time swap
- state update
- speaker swap
- near-duplicate paraphrase

Current dataset size:

- `573` cases