Information extraction
Recall the right fact from long histories.
BENCHMARK NOTEUPDATED 19 AUG 2026
UNBLOCK reports 93% on LongMemEval.
LongMemEval is an ICLR 2025 benchmark for long-term conversational memory. It tests whether a system can find, connect, update, time, and—when the evidence is missing—refuse to invent remembered facts.
Measured long-term memory
93% onLongMemEval.
UNBLOCK's reported answer accuracy on the 500-question benchmark for long-term conversational memory.
Read the benchmark note93%
0255075100
Information extraction Multi-session reasoning Knowledge updates Temporal reasoning Abstention
What LongMemEval measures
Memory is more than retrieval.
The official 500-question suite covers five abilities a useful memory layer needs in real conversations.
Multi-session reasoning
Combine evidence that appears across conversations.
Knowledge updates
Use the newest valid information when facts change.
Temporal reasoning
Understand when events happened and how they relate.
Abstention
Know when the memory does not contain a supported answer.
How to read the number
A result, not a magic trick.
Score: 93% answer accuracy, reported by UNBLOCK.
Benchmark: LongMemEval v1, published by Wu et al. and accepted at ICLR 2025.
Disclosure: The public run manifest, dataset revision, reader model, judge configuration, and per-question outputs are not attached to this release yet. Until they are, treat 93% as a company-reported result—not an independently verified leaderboard entry.
We publish that qualification because benchmark claims should be reproducible. This page will become the permanent home of the artifacts when the evaluation package is released.
The benchmark is proof. The product is live.