Fengzdadi / memory-bench-registry

# memory-bench-registry
**English** | 中文
**Browse the registry online →** Search and filter by memory type, reproducibility level and year.
A structured registry of agent memory benchmarks: one YAML file per benchmark (`benchmarks/ .yaml`). This README is generated by a script.
Every field must be traceable to the paper or the official repository; where each value came from is recorded in the entry's `sources`. Anything that cannot be found is `null`.
- Field and reproducibility-level definitions: docs/fields.md (in Chinese)
- Coverage-gap check against existing lists: docs/coverage-gaps.md (in Chinese)
- Existing curated lists and their verification status: docs/related-lists.md (in Chinese)
- Candidates reviewed but not included, with reasons: docs/excluded.md (in Chinese)
**207** entries. Reproducibility levels: `unverified` 0 · `paper_only` 37 · `data_public` 31 · `official_script` 139 · `smoke_tested` 0 · `full_run` 0
## By memory type
A benchmark can belong to several types. Expand a type to see its entries, newest first.
| Type | What the benchmark tests |
| --- | --- |
| `conversational_longterm` | Recall or reasoning over long, multi-session dialogue or chat histories (e.g. LoCoMo, LongMemEval, MSC) |
| `personalization` | Using user-specific preferences, personas or profiles to tailor answers (e.g. PersonaMem, PrefEval) |
| `episodic` | Recalling specific events and experiences with time, place or order ("what happened when") |
| `semantic_knowledge_update` | Factual knowledge stored outside the weights (context or external store), updated correctly when facts change |
| `procedural_workflow` | Distilling skills, workflows or tool use from earlier tasks and reusing them on *different* later tasks |
| `coding_cross_session` | Coding agents that must retain information across *independent* sessions or tasks |
| `parametric` | Memory in the model weights: knowledge editing, injection by fine-tuning, test-time training, memorization or unlearning |
| `multi_agent_group` | Memory shared among several agents, or multi-party / group conversations |
| `multimodal` | Remembered content includes non-text input (images, video, audio) that the system must process directly |
| `embodied` | Agents act in a simulated or real environment; memory is scored through actions or task success |
| `memory_security` | Attacks on or protection of memory: poisoning, injection, privacy leakage, whether deletion requests take effect |
Single-input long-context tests (e.g. RULER, NIAH, LongBench) are not memory benchmarks and are not included.
 conversational_longterm (81) 
| Name | Date | Links | Scale | Repro level |
| --- | --- | --- | --- | --- |
| DyadMem | 2026‑10 | paper | 3,065 conv. 61,210 Q | `paper_only` |
| RealCompanion | 2026‑10 | paper · code · data | 10 conv. 3,235 Q | `paper_only` |
| CUE-Mem | 2026‑09 | paper · code · data | 648 conv. 2,674 Q | `official_script` |
| DolphinBench | 2026‑09 | paper · code · data | 3 conv. 600 Q | `official_script` |