How LLMs Handle Memory: Key Techniques and Human Brain Parallels Explained

A technical discussion explores how large language models manage memory, noting that no model has a truly infinite context window and that continuous dialogue is simulated through compression and selection. Several architectural approaches exist to extend effective memory, including Google's Infini-attention, StreamingLLM's sliding window method, MemGPT's three-tier virtual memory system, and Mem0's selective fact storage, which can cut token usage by 80–90%. The piece also draws comparisons to human memory, highlighting that the brain reconstructs rather than replays information — a principle first demonstrated by Bartlett in the 1930s — and that forgetting is an active consolidation process, not mere data loss. A notable practical concern is raised around tokenization: processing Russian text costs roughly 70% more tokens than English due to its rich inflectional morphology, diluting BPE token efficiency. Research by MIT's Evelina Fedorenko further suggests that the brain's language network is largely separate from systems handling logic, math, and social reasoning, challenging assumptions about the relationship between language and thought.