Memory
Engram
2024ResearchUpdated: 15 September 2026Published
Key
innovation
A concept borrowed from neuroscience of a persistent memory trace, applied to persistent-memory mechanisms in AI models and agents.
Category
Memory
Abstraction level
Building block
Operation level
Architecture blockInference
Use cases
Long-term memory for AI agentsPersonalization and interaction continuityMemory and knowledge systems
How it works
Information is encoded into persistent representations (e.g. vectors, entries in external memory or a knowledge base), stored outside the active context window and recalled through retrieval or associative addressing when relevant to the current task.
Problem solved
Models and agents lose context beyond the context window. Engrams provide a way to persistently store and recall memory across sessions.