Evaluation · Benchmark
The LongMemEval Benchmark Explained
LongMemEval stress-tests long-term memory across sessions — measuring whether agents retain and retrieve facts after gaps between conversations — and is the headline benchmark cited for Engram, Mem0, Zep and competitors.
LongMemEval tests
Definition
What LongMemEval measures
Cross-session recall, multi-hop memory and temporal gaps between conversations.
LongMemEval tests whether agents remember facts from prior sessions after time gaps — the stress test for durable long-term memory. Tasks include single-hop recall, multi-hop reasoning across stored facts and temporal reasoning (“what was true when?”).
Results
Scores and published results
Last updated: July 2026. N/A = no public result.
| Framework | LongMemEval | Source |
|---|---|---|
| Engram | N/A | GA June 2026 — pending |
| Zep | 71.2% (gpt-4o), +18.5% vs baseline | Rasmussen et al., 2025 |
| Mem0 | N/A (no public result) | — |
| LangMem | N/A (no public result) | — |
Comparison
LongMemEval vs LOCOMO
LongMemEval = cross-session gaps; LOCOMO = long context within one chat.
Complementary benchmarks. Run both before production: LOCOMO for in-session depth, LongMemEval for multi-session persistence. See LOCOMO benchmark.
Transparency
Why we cite LongMemEval in comparisons
LongMemEval is the most widely referenced cross-session benchmark in the agent-memory space.
Our compare cluster cites published LongMemEval numbers with sources and dates. We never estimate unpublished scores — Engram rows show N/A until results are published.
FAQ
Frequently asked questions
LongMemEval vs LOCOMO?
LongMemEval tests cross-session recall with temporal gaps. LOCOMO tests long single-conversation recall. Use both. See LOCOMO benchmark.
What is Mem0's LongMemEval score?
N/A — no public LongMemEval result published. Mem0 LOCOMO J 66.9 (Chhikara et al., 2025).
What is Zep's LongMemEval score?
71.2% accuracy with gpt-4o, +18.5% vs baseline (Rasmussen et al., 2025). DMR 94.8%.
What is Engram's LongMemEval score?
N/A publicly as of July 2026. Engram reached GA June 2026. See Engram explained.
Is LongMemEval an open dataset?
LongMemEval is a published benchmark used in agent-memory research. Check the original paper for dataset access.
Can you run LongMemEval locally?
Yes — integrate your memory framework, run the eval harness and compare against published baselines. See add memory guide.
Where is the LongMemEval leaderboard?
Scores are published in framework papers — Zep (Rasmussen et al., 2025) is the primary LongMemEval citation. We track results on best AI memory tools.
LongMemEval updates in 2026?
We update published scores as frameworks release new results. Engram GA June 2026 — benchmark rows pending.