Long way to go with AI persistent memory
retained text
A state management problem. We need a system that can extract facts events from conversations, track importance and then retrieve them based on both semantic relevance and time. Then you need mechanisms for deduplication, contradiction handling, consolidation and decay. probably the critical problem is deciding what enters the context. Can't keep injecting thousands of memories into every inference can I, that's why you need retrieval, re ranking and context compression to select the most useful subset. and that means maintaining an evolving state of what the model knows, updating it when things change and knowing what information is worth carrying into the next interaction. RAG can retrieve external information, but that's still fundamentally a retrieval game. A memory system needs to maintain a persistent state outside the context window. The system needs to maintain a persistent representation of its current knowledge, continuously update that state as new information arrives and that's fundamentally different from simply expanding the context window.
document fetched 2026-09-12 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wdmi0e/long_way_to_go_with_ai_persistent_memory