| # Design Notes |
|
|
| ## Purpose |
|
|
| Most memory benchmarks measure semantic recall in benign settings. |
|
|
| This benchmark targets retrieval failure modes that matter in agent memory systems: |
|
|
| - retrieving the wrong person |
| - retrieving the wrong environment |
| - retrieving an outdated fact instead of the current one |
| - retrieving something semantically close but operationally wrong |
|
|
| ## Benchmark principles |
|
|
| - retrieval-focused, not generation-focused |
| - one query, many plausible distractors |
| - exact relevant entry ids are known in advance |
| - metadata such as tags, depth, speaker, and timestamp may be present but are optional |
| - cases should remain small enough to inspect by hand |
|
|
| ## Scoring |
|
|
| Suggested scoring: |
|
|
| - `Recall@1` |
| - `Recall@5` |
| - `MRR` |
| - error bucket counts by `adversary_type` |
|
|
| ## Expansion ideas |
|
|
| - more software-specific adversaries |
| - benchmark splits by domain |
| - fact-update and contradiction-specific suites |
| - Hugging Face dataset packaging |
|
|
