Benchmarks
Memory accuracy, measured
Hindsight leads every long-term memory benchmark with a published comparison, from 94.6% on LongMemEval-S to BEAM at 10 million tokens. The research is peer-reviewed and the benchmark code is open source.
Retrieval accuracy
Hindsight against the next-best system on each dataset. The runner-up is a different system on almost every row, so it isn’t named; where nobody else has published, the row says so.
- LongMemEval-S
- 94.6%74.0%
- LoComo
- 92.0%80.3%
- PersonaMem
- 86.6%84.4%
- PrecisionMemBench
- 85.7%
no published comparison
- LifeBench
- 71.5%61.0%
- BEAM · 10M tokens
- 64.1%40.6%
Coding agents
On AMB’s sdebench, a coding agent works through 61 real repository tasks, three runs each, with and without memory. Accuracy doesn’t move: every run solves 60–61 of the 61 either way. What memory changes is how often you have to stop and correct the agent, and what the tokens cost.
| Agent | Corrections / task | Cost / task |
|---|---|---|
| Claude Code | 0.85 → 0.36 | $0.445 → $0.338 |
| Codex CLI | 1.34 → 0.47 | $0.577 → $0.276 |
| opencode | 1.20 → 0.80 | $0.634 → $0.553 |
No memory → with Hindsight, mean of 3 runs.
Methodology and sources
- All benchmark results
- Every dataset, configuration and run.
- Agent Memory Benchmark
- The live leaderboard behind most of the comparisons above.
- Read the paper
- The peer-reviewed research behind Hindsight.
Start building agents that learn
Open source, MIT licensed. Self-host or use Hindsight Cloud.
94.6%
LongMemEval-S retrieval accuracy · next best published system 74.0%