Microsoft Research Unveils Memora, a Long-Term Memory Architecture for AI Agents
Memora addresses a core bottleneck for AI agents working over long periods by enabling them to retain detailed context without overwhelming token counts or losing specifics in summaries.
Reporting from 1 sources: GIGAZINE.
Microsoft Research released Memora, a long-term memory architecture for AI agents that balances storing fine details with efficient retrieval. It separates stored content from search entry points, using primary abstractions and cue anchors to allow agents to recall information from multiple clues, and includes a policy-guided retriever that revises search strategies when initial results are insufficient.
Microsoft Research published Memora on June 29, 2026, a memory architecture designed to let AI agents store conversation and task histories over long periods and retrieve only what is needed. The system separates stored content, called memory values, from search entry points: primary abstractions summarize the subject, while cue anchors provide alternative paths via related names, schedules, or topics. A policy-guided retriever iteratively refines the search strategy, expanding clues when information is lacking and stopping once enough is gathered. Memora was tested on the LoCoMo benchmark for long conversations.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.