AI Agents Engineering · What an Agent Actually Is · Lesson 3 of 3
Planning & memory
Two capabilities turn a bare loop into a competent agent: planning (decide the shape of the work before acting) and memory (carry information across steps and sessions).
Planning
- Decompose first: for complex tasks, have the model draft a plan / sub-goals, then execute — this reduces flailing and gives you something to check against.
- Reflection: let the agent critique its own output or a step result and retry — a cheap reliability win.
- Plans can be static (made once) or replanned as new observations arrive (plan-and-execute).
Memory
- Short-term = the running context window (recent thoughts, tool results). Finite and expensive.
- Long-term = an external store (vector DB, key-value scratchpad, files) read and written across steps and sessions.
- Working memory hygiene: summarize or prune old turns so the context does not blow the window or bury the goal.
RAM vs diskShort-term memory is RAM — fast, small, wiped between runs. Long-term memory is disk — slower to reach, but durable across sessions. Good agents move the right things between them.
◆ Lock it in
- Plan → execute → reflect beats a naive loop on complex tasks.
- Memory splits into short-term (context) and long-term (external store).
- Manage the context actively — summarize/prune so the goal is not buried.
Feynman drill — say it out loudExplain the difference between short-term and long-term agent memory, and when you would write to each.
Step 1 rate your confidence · Step 2 pick your answer
What best describes 'long-term memory' for an agent?
Step 1 — how sure are you?