CHAPTER 03 / 7 MIN READ
Remember deliberately
Memory is a policy for selecting useful state, not a pile of old text.
START HERE
At the beginning of a delivery task, the agent receives the door code. Many observations later, it reaches the door. The code is still in the log—but no longer in the model’s active input.
Stored information helps only when the harness retrieves it and makes it available at the right moment.
1.Separate working context from storage.
Working context is what the current model call can consume. External storage can be much larger, but it is not automatically visible. A memory system needs a way to write, address, retrieve, and reintegrate information.
MemGPT uses a virtual-memory analogy: maintain a limited working set and move information between tiers. The agent can use functions to manage parts of its own memory rather than relying entirely on an ever-growing conversation.
2.Retention and retrieval solve different problems.
Retention asks what should be saved. Retrieval asks what is useful now. A complete archive with poor retrieval can still leave the agent unable to answer. A compact summary can be useful but may erase the exact detail the next task needs.
A practical memory entry benefits from context: what it means, where it came from, and whether it might be stale. “4821” is weaker than “door code for the east entrance, supplied for this delivery.”
3.A bigger window postpones the decision.
Increasing the context capacity can keep more observations available. It does not make every observation equally relevant, and a long enough task will still require selection.
The corpus distinguishes an append-only trajectory from an addressable memory store. That is why its profile can mark an agent loop as present while marking memory as absent. It is a definition used for comparison, not a claim that the model has forgotten everything.
The delivery code
BEFORE
Keep only the newest three observations. The early door code falls out of view.
AFTER
Save the code separately. At the door, retrieve the entry and place it in the active context.
The storage did not become the model’s memory by magic. A retrieval step connected it to the next decision.
PAUSE & PREDICT
A useful fact is stored on disk. Why might the agent still miss it?
The toy experiment assumes perfect retrieval. Real memory systems can retrieve irrelevant, stale, or conflicting entries and must manage a finite input budget.
GO TO THE SOURCE
MemGPT2023 ↗This chapter is an original teaching synthesis. Open each research note for source links, findings, and classification caveats.
TAKE IT INTO YOUR OWN WORK
Choose one recurring task. Name a fact to retain, a retrieval cue, and a condition that would make that fact stale.