CH-Bench.

A zero-dependency Python evaluation harness for retrieval and answer quality, with adapters for different memory systems.

CH-Bench measures persistent-memory systems through a common adapter instead of presenting an impressive graph as evidence. The selected résumé and local documentation describe retrieval metrics such as recall@k, MRR and nDCG, answer evaluation for correctness, groundedness and abstention, and efficiency reporting. The harness supports LongMemEval, LoCoMo and custom long-horizon recall tasks. These are available evaluation capabilities, not a claim of a winning benchmark result. Its adapter boundary lets a memory implementation be compared using the same task and reporting workflow.

• 01

  Retrieval and answer evaluation are separate

• 02

  Four-method system adapter

• 03

  Groundedness and abstention checks

• 04

  Zero-dependency core