Cognition, Planning & Memory: NCP-AAI Domain 5 (10%)

What this domain covers

Cognition, Planning, and Memory is about how an agent thinks: how it turns a goal into a sequence of steps, decides what to do next, and carries information forward across a long task. It is 10% of the exam — the heaviest domain after the top four — because planning and memory are what separate a multi-step agent from a single prompt.

Planning and task decomposition

Hard tasks are solved by breaking them into smaller sub-tasks. You should know the main planning strategies: ReAct interleaves reasoning and acting one step at a time; plan-and-execute drafts the whole plan up front and then carries it out; and decomposition splits a goal into sub-goals handled in sequence or by sub-agents. Up-front planning is more predictable; step-by-step planning adapts better to surprises.

ReAct:            think -> act -> observe -> think -> ...  (interleaved)
plan-and-execute: draft full plan -> run each step -> replan if needed
decomposition:    goal -> [sub-goal, sub-goal, ...] -> combine

Memory: working, short-term, and long-term

Agents use several kinds of memory. Working (short-term) memory is the live context window — the running conversation and recent tool results. Long-term memory persists across sessions in an external store, often a vector database, and is retrieved when relevant. People also distinguish episodic memory (what happened — past interactions) from semantic memory (durable facts the agent knows). The skill is deciding what to keep in-context versus what to offload and retrieve.

Working / short-term:  the context window — recent turns + tool results
Long-term (episodic):  past interactions, stored and retrieved by relevance
Long-term (semantic):  durable facts / knowledge about the user or domain

Managing the context window

The context window is finite, so long tasks must manage it actively. Common techniques: summarize older turns into a compact running summary, retrieve only the memories relevant to the current step instead of stuffing everything in, and drop or compress stale tool output. Letting context grow unbounded raises cost and latency and eventually overflows the window.

Reflection and self-correction

Stronger agents review their own progress: checking whether a step succeeded, noticing when they are stuck or looping, and revising the plan. Reflection — having the agent critique its own trajectory and adjust — improves reliability on hard tasks at the cost of extra model calls. It pairs naturally with memory: the agent records what worked and avoids repeating mistakes.

Exam tip

When a scenario describes an agent that "forgets" earlier steps or runs out of context on a long task, the answer is usually a memory/context strategy — summarization, external long-term memory, or relevance-based retrieval — not simply a bigger context window or a bigger model. And match planning style (up-front vs step-by-step) to how dynamic the task is.