// MEMORYGRAPH_VS_MEM0_
> Benchmark Comparison: Experience Reuse Performance
// GOOGLE_DEEPMIND_EVO_MEMORY_BENCHMARK
The Evo-Memory benchmark measures how well AI memory systems enable "experience reuse" — learning from past sessions to improve future performance. This is different from simple recall.
| Benchmark | Mem0 | No Memory | Delta |
|---|---|---|---|
| AlfWorld | 0.27 | 0.28 | -0.01 |
| ScienceWorld | 0.32 | 0.31 | +0.01 |
| PDDL | 0.10 | 0.08 | +0.02 |
| BabyAI | 0.54 | 0.52 | +0.02 |
Mem0 provides ≤2% improvement over no memory at all.
Source: Evo-Memory: Evolving Memory Mechanisms for Agentic Systems, arXiv:2511.20857v1 (Dec 2025)
// WHY_THE_DIFFERENCE
"Agents remember what was said but not what was learned."
Mem0 uses flat vector similarity:
- No understanding of relationships between memories
- Retrieves facts, not strategies
- Cannot traverse "Problem → Solution" chains
MemoryGraph uses typed graph relationships:
- SOLVES links connect problems to solutions
- CAUSES links explain why things fail
- IMPROVES links track evolution of approaches
- Graph traversal retrieves relevant experience, not just text
// FEATURE_COMPARISON
| Feature | MemoryGraph | Mem0 |
|---|---|---|
| Storage Model | Graph (35+ relationship types) | Flat + Vector |
| Relationship Tracking | [✓] Native | [✗] None |
| Problem→Solution Links | [✓] SOLVES relationship | [✗] Not supported |
| Memory Consolidation | [✓] Dreams Agent | [✗] None |
| Local-First | [✓] SQLite default | [✓] Local option |
| MCP Tools | 9-11 | 4-5 |
| License | Apache 2.0 | Apache 2.0 |
// READY_TO_UPGRADE
Experience the power of graph-based memory for your AI assistants.