Memory benchmarks
How Smart Jelly’s memory works.
What it remembers, how well it recalls it, and how we measured that, with every answer published so you can re-score it.
87.6%BEAM-100K · 400 questions
Second of all the memory apps that published a score: behind past.dev (91.8%), ahead of Exabase (76.9%), Hindsight (75.0%) and Honcho (63.0%).
One memory of your conversations.
It reads only the apps you choose. Everything it remembers links back to where it came from.
Your daysWhat happened each day: chats, calls and work.Tue 4:00 PM · Google Meet with Stripe partnerships · 25 min
People and factsWho people are and what you’re working on, with where each fact came from.Stripe partnerships · usage-meter spec · 3 conversations
How you workYour habits and how you like things done. You can fix or remove any of them.Short replies, no sign-off
Finding answersFinds the right memories when you ask, and checks the answer.“When is my 83(b) deadline?” · with its source
About the BEAM benchmark
Published BEAM-100K results.
BEAM is a public test: hundreds of questions about long conversations. Each score is that team’s own. Only past.dev ran it the same way we did, so compare the others loosely.
| System | Score | AI model used | Source |
|---|---|---|---|
| past.dev | 91.84% | AI model usedgpt-5.6-luna | Sourcepast.dev/benchmarks · 16 and 19 Sep 2026 |
| Smart Jelly | 87.57% | AI model usedgpt-5.6-luna | Sourcethis page · run of 29 Sep 2026 |
| Exabase M-1 | 76.9% | AI model usedGemini 3 Flash | Sourceexabase.io · 19 Sep 2026 |
| Hindsight | 75.0% | AI model usednot stated | Sourcebenchmarks.hindsight.vectorize.io |
| Honcho | 63.0% | AI model usedClaude Haiku 4.5 | Sourceplasticlabs.ai · 19 Dec 2025 |
| LIGHT (BEAM paper) | 35.8% | AI model usedLlama-4-Maverick | SourceTavakoli et al., ICLR 2026 |
| RAG baseline (BEAM paper) | 32.3% | AI model usedLlama-4-Maverick | SourceTavakoli et al., ICLR 2026 |
Checked against each source on 29 September 2026.
How we got 87.6%.
The testAll 400 questions of BEAM-100K, across ten abilities, over long conversations.
Answering and judginggpt-5.6-luna with reasoning off for both, the same as past.dev.
ScoringEach question scores 0 to 1; the headline is the mean over all 400.
Every answer, public.
The results repository has every question, answer and judge verdict, and a script that recomputes 87.57% with no API calls.
See it on GitHub