Mem0 Research Paper: Token-Efficient Memory Algorithm
Mem0 Research: Benchmarking a Token-Efficient Memory Algorithm for AI Agents
Benchmarked across LoCoMo, LongMemEval, and BEAM Powered by single-pass hierarchical extraction and multi-signal retrieval
Summary
Mem0's new token-efficient memory algorithm hits 92.5 on LoCoMo, 94.4 on LongMemEval, and 64.1/48.6 on BEAM (1M/10M) while averaging under 7,000 tokens per retrieval call. Full-context approaches on the same benchmarks use 25,000+. High accuracy at 3-4x lower token cost.
BENCHMARKS
LoCoMo
1,540 questions • 5 categories
92.5
OVERALL
6956
Mean Tokens
| Single-hop | Multi-hop | Open-domain | Temporal | Old | New |
|---|---|---|---|---|---|
| 76.6 | 94.6 | 70.2 | 95.4 | 57.3 | 82.3 |
| 63.2 | 92.5 |
LongMemEval
500 questions • 6 categories
94.4
OVERALL
6787
Mean Tokens
| Single-session (user) | Single-session (assistant) | Single-session (preference) | Knowledge update | Temporal reasoning | Multi-session | Old | New |
|---|---|---|---|---|---|---|---|
| 94.3 | 98.6 | 46.4 | 98.2 | 76.7 | 96.7 | 79.5 | 93.6 |
| 51.1 | 97.0 | 70.7 | 88.0 |
BEAM
BEAM 1M: 700 questions • 35 conversations
BEAM 10M: 200 questions • 10 conversations
64.1
OVERALL (1M)
48.6
OVERALL (10M)
6719
Mean Tokens (1m)
6914
Mean Tokens (10m)
| Preference Following | Instruction Following | Information Extraction | Knowledge Update | Multi Session Reasoning | Summarization | Temporal Reasoning | Event Ordering | Abstention | Contradiction Resolution |
|---|---|---|---|---|---|---|---|---|---|
| 88.3 | 90.4 | 85.2 | 82.5 | 70.0 | 56.3 | 65.0 | 75.0 | 65.2 | 26.1 |
| 63.5 | 46.9 | 61.8 | 16.3 | 53.6 | 20.2 | 52.5 | 40.0 | 35.7 | 32.5 |
Data last updated: May 2026.
All results are Old Algorithm vs. New Algorithm.
Full evaluation framework is open-sourced on GitHub.
WHAT’S NEW
Single pass ADD-only extraction
Mem0 now treats agent-generated facts as first-class, closing a significant gap in memory coverage. When an agent confirms an action or provides a recommendation, that information is stored with equal weight.
Multi-signal retrieval
Retrieval stack now runs three scoring passes in parallel and fuses the results: Semantic similarity, Keyword matching, and Entity matching. The combined score outperforms individual signal scores.
What we're building next
Temporal abstraction
Representing how events relate over time, not just what happened. BEAM 10M scores define the current frontier.
Cross-session structure
Modeling how information evolves across sessions. Requires connecting scattered interactions into coherent timelines.
Agent-native memory
Extraction and retrieval running asynchronously as infrastructure, so agents don’t spend cycles managing their own context.