Back to Arena
MemGPT Classic
by Berkeley / Letta
System Card
OrganizationBerkeley / Letta
Released2023-10
Architectureagentic-workflow / Virtual context with paging
DetailsOriginal MemGPT formulation: an LLM treats its context window as a paged virtual memory, calling tools to swap data between in-context working memory and an external archival store.
Parameters—
Domainagent-memorylong-contextepisodic-session
Open SourceYes
PaperView Paper
CodeRepository
memgptvirtual-contextresearchpaging
Capability Profile
Benchmark Scores
3 of 14 benchmarksLong-Context Retrieval0/5
RULER
no dataNIAH
no dataLooGLE
no dataLongBench
no data∞Bench
no dataMulti-Turn Recall1/2
MemoryBank
no dataCross-Session Memory1/1
Multi-Hop QA0/3
BABILong
no dataMultiHop-RAG
no dataHotpotQA
no dataAgent Task Memory1/1
Personalization0/1
PerLTQA
no dataFactuality / Grounding0/1
RAGAS
no data