Memory meets context

Compaction is short-term cleanup. Memory is long-term storage. They are not the same layer. Think of compaction as RAM management and memory as the hard drive. The model can survive losing detail in RAM as long as the important facts are on disk.

Two layers of durability

Conversation flows into history (RAM). Important facts flow from the agent into memory (disk). Compaction trims history. Memory persists.

Imagine you tell the agent your laptop is a MacBook Air M2 on Monday. By Friday the Monday session has been compacted away. On Friday you ask 'what version of Python should I install?'. Without memory, the agent has no idea you are on Apple Silicon. With memory, the agent's first move is memory_search('user setup'), which returns 'MacBook Air M2', and now the answer is grounded.

Validation checklist: Build it: combine compaction and memory

Loading practice…

Quiz: Quiz

Loading practice…