Transcript note: a short passage (on patient H.M.’s preserved abilities) came out garbled; the surrounding context is intact.
Neuroscientist meytar-zemer (Weizmann Institute; now head of research at a medical unstructured-data company) compares human memory with AI memory. Human memory is encoded, stored, and retrieved. It splits into short-term/working memory and long-term memory (episodic, semantic, procedural), as shown by the amnesic patient H.M. The hippocampus binds experiences and talks with the cortex; replay, mostly during sleep, strengthens synapses until memories rely on the cortex; the amygdala tags emotional valence. Retrieval reassembles a memory from a cue and makes it editable again (reconsolidation), so even “flashbulb” memories like 9/11 drift. AI contrast: vector databases are unlimited and static; Mem0-style memory extracts facts with a generic prompt unaware of task context. Humans excel at separating signal from noise, tagging importance, and organizing knowledge into schemas, and at continuous learning, all still missing in AI.
Show: langtalks
Concepts: human-vs-ai-memory · memory-consolidation · autonomous-agents-outlook · knowledge-graphs