AI memory performance proven to be top-tier
Memory.Inc Achieves 94.8% on LongMemEval-S
Memory.Inc scored 94.8% on LongMemEval-S, a leading AI memory performance benchmark.
This is pre-launch performance, matching SOTA compared to major public AI memory systems.
LongMemEval-S is a top benchmark checking how well AI recalls key details from long chats and replies using scattered context.
Memory isn't just finding answers in one file like simple search tests. It must combine user details, past replies, updated facts, user tastes, timelines, and relative dates across multiple chat sessions.
Simply put, it doesn't just test "finding similar sentences" but tests key memory skills needed when AI chats with real users long-term.
Category | Memory.Inc | Mastra OM | Supermemory | Zep | Full Context |
|---|---|---|---|---|---|
single-session-user | 95.7% | 98.6% | 97.1% | 92.9% | 81.4% |
single-session-assistant | 100.0% | 82.1% | 96.4% | 80.4% | 94.6% |
single-session-preference | 96.7% | 73.3% | 70.0% | 56.7% | 20.0% |
knowledge-update | 97.4% | 85.9% | 88.5% | 83.3% | 78.2% |
temporal-reasoning | 95.5% | 85.7% | 76.7% | 62.4% | 45.1% |
multi-session | 83.5% | 79.7% | 71.4% | 57.9% | 44.3% |
Total | 94.8% | 84.23% | 81.6% | 71.2% | 60.2% |
Scroll left and right to view the entire table.
Supermemory's public 95% score is based on Recall@15 with aggregation, using multiple search results. This table uses basic LongMemEval-S QA accuracy for a fair comparison.
MemKraft, MemPalace, etc., are excluded as they cover only some sub-criteria or different tests instead of all 500 LongMemEval-S items.
The real challenge in AI memory isn't just "storing lots of data."
It is about exact recall.
And even more, keeping updated details fresh.
If a user said A but changed it to B later, the AI shouldn't rely on A. It must answer based on the new Info B.
AI must adapt as budgets, tastes, schedules, and project goals change.
Memory.Inc is not a simple chat log store, but an AI memory system built to keep user and team contexts dynamically updated.
Post-launch, we will open-source our evaluation code so users can run and verify the benchmark themselves.
Memory.Inc is building the memory infrastructure for AI to recall accurately and keep context longer.