planning and memory in agents
/ PLAN-ing and MEM-uh-ree in AY-junts /
Planning is an agent breaking a big goal into an ordered set of smaller steps before diving in; memory is its way of holding on to what it has learned or done so it isn't starting from scratch every moment. The analogy is a traveler with a checklist and a notebook: the checklist (plan) says what to do and in what order; the notebook (memory) records what happened, what worked, and what to remember for later. Without both, an agent lurches blindly and forgets its own footsteps.
Why these need special handling: the model at an agent's core has a fixed-size "working memory" — its context window — that can only hold so much text at once, and it carries nothing over between separate sessions on its own. So planning is added by prompting the model to lay out steps, or by code that tracks a to-do list; memory is bolted on with outside storage — a running summary of the conversation, a scratchpad of intermediate results, or a database the agent can search to recall facts from long ago (often using the same retrieval techniques as retrieval-augmented generation).
Be honest about how shaky this still is. The model doesn't truly "remember" — it re-reads stored text that the surrounding system chooses to show it, and if the system forgets to retrieve the right note, the agent forgets too. Plans made up front can become stale the moment reality differs from expectation, and a brittle agent will keep marching through a plan that no longer makes sense. Good planning and memory are as much careful engineering of what to store and when to recall it as they are anything the model does by itself.
An agent told to "plan a three-day trip" first writes a plan: book flights, then hotel, then activities. As it works, it jots notes to memory — "user prefers window seats," "hotel booked, confirmation #X." A week later, asked to "add a dinner reservation," it searches that memory, recalls the dates and hotel, and slots the dinner in without re-asking everything.
Plan = ordered steps; memory = stored notes the system feeds back when relevant.
An agent's "memory" is not the model remembering — it is external storage that the surrounding software retrieves and pastes back into the prompt. Nothing persists inside the model between calls. So memory is only as good as the retrieval: the wrong note recalled (or the right one missed) leads the agent astray just as surely as forgetting.