| --- |
| license: apache-2.0 |
| task_categories: |
| - question-answering |
| language: |
| - en |
| tags: |
| - benchmark |
| - memory |
| - ai-agents |
| - longitudinal |
| - evaluation |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # MemoryStress Dataset |
|
|
| **The first longitudinal benchmark for AI memory systems.** |
|
|
| 583 facts x 1,000 sessions x 300 questions x 6 scoring dimensions. |
|
|
| **Website:** [omegamax.co](https://omegamax.co) | **Memory System:** [github.com/omega-memory/core](https://github.com/omega-memory/core) | **Benchmark Code:** [github.com/omega-memory/memorystress](https://github.com/omega-memory/memorystress) |
|
|
| ## Overview |
|
|
| Every existing memory benchmark tests recall from a handful of sessions. MemoryStress tests what happens at session 1,000 — when facts contradict, memories fade, and noise accumulates over 10 simulated months. |
|
|
| ## Dataset Structure |
|
|
| ```json |
| { |
| "version": "1.0", |
| "phases": { "1": {...}, "2": {...}, "3": {...} }, |
| "stats": { "total_facts": 583, "total_sessions": 1000, ... }, |
| "facts": [ { "fact_id": "F001", "content": "...", ... } ], |
| "contradiction_chains": [ { "chain_id": "C001", ... } ], |
| "sessions": [ { "session_id": "S0001", "turns": [...], ... } ], |
| "questions": [ { "question_id": "Q001", "question": "...", ... } ] |
| } |
| ``` |
|
|
| ## Key Features |
|
|
| - **583 facts** across 6 categories (preferences, decisions, technical, personal, events, relationships) |
| - **1,000 sessions** across 3 phases of increasing noise |
| - **40 contradiction chains** where facts update, revert, accumulate, or partially change |
| - **300 questions** at 4 phase checkpoints, spanning 7 question types |
| - **6 scoring dimensions**: recall@age, degradation curve, contradiction resolution, cost efficiency, cross-agent recall, cold start recovery |
|
|
| ## Three Phases of Pressure |
|
|
| | Phase | Sessions | Description | |
| |-------|----------|-------------| |
| | Phase 1 | 1-100 | Low noise, establishing baseline | |
| | Phase 2 | 101-500 | Growing noise, contradictions emerge | |
| | Phase 3 | 501-1000 | Dense, high-entropy, multi-topic stress | |
| | Phase 4 | Recovery | No new sessions — can you still recall? | |
|
|
| ## Usage |
|
|
| ```bash |
| pip install memorystress |
| |
| # Run the null baseline (free, instant) |
| memorystress run --dataset memorystress_v1.json --adapter null --grade |
| |
| # Run with OMEGA |
| memorystress run --dataset memorystress_v1.json --adapter omega --grade --extract-facts |
| ``` |
|
|
| ## Benchmark Code |
|
|
| Full benchmark framework: [github.com/omega-memory/memorystress](https://github.com/omega-memory/memorystress) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @software{memorystress2026, |
| title={MemoryStress: The First Longitudinal Benchmark for AI Memory Systems}, |
| author={OMEGA Memory}, |
| url={https://github.com/omega-memory/memorystress}, |
| year={2026} |
| } |
| ``` |
|
|
| ## License |
|
|
| Apache-2.0 |
|
|
