Push vs Pull Memory: A Better Way to Think About AI Agent Memory

Short summary

Push and pull memory represent two architectures for AI agent systems: pull queries a passive store at read time, leaving contradiction reconciliation to readers; push has agents read before acting and write corrections back, with the substrate automatically superseding stale facts. Pull works for static facts and short sessions, while push suits long-lived agents in changing contexts. This is now practical because LLM agents can reliably handle the read-correct-write loop.

• •Push memory moves reconciliation from read time to write time, automatically superseding stale facts
• •Pull memory suits static facts and short sessions; push suits long-lived agents where correctness matters as the world changes
• •Push is practical now because LLM agents can reliably read current state, act, and write structured corrections

Generated with AI, which can make mistakes.