Benchmark
Benchmark
deep-dives
AI agent memory benchmark results across LoCoMo, LongMemEval, and BEAM
LoCoMo
LongMemEval
BEAM (1M)
BEAM (10M)
LoCoMo
1,540 questions 5 categories
Tests single-hop, multi-hop, open-domain,
and temporal memory recall across multi-session conversations.
92.5
Mem0 Score
6,956
Mean Tokens
Mean tokens: 6,956. Biggest gains on temporal (+29.3) and multi-hop (+25.2).
Single Hop
94.6
Multi-Hop
95.4
Open-domain Temporal
82.3
Temporal
92.5
Methodology
Why the numbers
look this way?
Each score above traces back to a specific piece of the architecture and not just asserted.
Recall Correctness
The ADD-only extraction never overwrites a fact, entity linking ties together mentions of the same person, place, or concept, and temporal reasoning reranks toward what's current. The right memory surfaces on the first retrieval call.
Context Footprint
Three retrieval signals, semantic, keyword, and entity, are scored in parallel and fused into one ranked result. Only the top matches enter the prompt, rather than the full conversation history, which is what keeps each call near 7,000 tokens instead of 25,000+.
Response Time
Memory decay's recency ranking runs fire-and-forget, and temporal reasoning does its classification at write time, so median latency stays flat at +1ms.
New
Blog
Every number on this page comes from a benchmark harness anyone can run, audit, or extend dataset loaders, judge-model config, and the exact run configs behind these results.
