MemoryBench
A benchmark for evaluating the memory capabilities of AI models.
Core Facts
Timeline (last 90 days)
Meta's MemoryBench checks whether a model can update facts within a single long conversation, but explicitly does not test persistent memory across sessions.
MemoryBench is a benchmark created by Meta itself, testing recall of key information in multi-turn long conversations of 64K to 100K tokens, with conversations spanning multiple facts and including distractor conversations as well as injected turns that conflict with or update the target details, scored by an LLM judge according to weighted scoring criteria.
In the MemoryBench table, the scores of the five models are 85.5, 85.6, 79.0, 81.4, and 81.5 respectively, with a gap of only 0.1 points at the top.
In the MemoryBench test, Muse Wonder achieved 85.5% accuracy, exceeding Claude Opus 5.5's 81.5% and GPT-6 Astra's 81.4%.
On MemoryBench, Wonder defeated Fable, Opus, and Astra; on the Artificial Analysis index, Wonder Full ranks second, behind only Opus 5.5.
MemoryBench and the 'Personal Superintelligence Index' are both Meta internal benchmarks using LLM judging, which carry vendor self-evaluation bias.
Multi-turn conversation configuration addresses the context memory problem, determining how many rounds of historical information an agent can retain during continuous dialogue.
All Facts (4)
Multi-turn conversation configuration addresses the context memory problem, determining how many rounds of historical information an agent can retain during continuous dialogue.
75%UnverifiedMeta's MemoryBench checks whether a model can update facts within a single long conversation, but explicitly does not test persistent memory across sessions.
50%UnverifiedMemoryBench is a benchmark created by Meta itself, testing recall of key information in multi-turn long conversations of 64K to 100K tokens, with conversations spanning multiple facts and including distractor conversations as well as injected turns that conflict with or update the target details, scored by an LLM judge according to weighted scoring criteria.
50%UnverifiedMemoryBench and the 'Personal Superintelligence Index' are both Meta internal benchmarks using LLM judging, which carry vendor self-evaluation bias.
50%