Module 2 of 10, lesson 1 of 5

Anatomy: components and memory

Name the modules of an agent and choose the types of memory it actually needs.

You can describe an agent as a cycle of modules. Each depends on the one before: a mistake at the input spreads to every later decision.

1. Perception: receive and clean inputs (requests, APIs, logs, sensors).
2. Reasoning: weigh the options.
3. Planning: split the goal into steps and dependencies.
4. Action: execute, often through a tool call.
5. Communication: report to a human or another agent.
6. Learning: use the result to do better next time.

One caveat. In most current LLM agents, these modules aren’t separate components: reasoning and planning happen in the same model call. The breakdown helps you think about the system, not necessarily code it.

Four memories

“Memory” covers very different needs.

• Working memory: the context visible right now. It has a ceiling, and performance degrades when you overfill it.
• Semantic memory: what the agent knows in general. In practice, this is often plain Markdown files loaded at startup.
• Procedural memory: how to do things. A short description stays visible, and the detailed instructions load only when needed.
• Episodic memory: what happened. A raw transcript helps little; a summary of decisions and their lessons helps a lot.

Forgetting is a feature

Memory that only grows ends up polluting decisions. You have to decide what expires, what gets summarized and what gets corrected. With personal data, this is also a legal duty.

Add a memory to solve a problem you’ve observed, not because the diagram shows four.

function pickMemories(memories, budget) {
  const sorted = [...memories].sort((a, b) => b.score - a.score);
  const kept = [];
  let used = 0;
  for (const memory of sorted) {
    if (used + memory.tokens > budget) continue;
    kept.push(memory.id);
    used += memory.tokens;
  }
  return kept;
}

const memories = [
  { id: 'refund-policy', score: 0.9, tokens: 300 },
  { id: 'full-transcript', score: 0.8, tokens: 900 },
  { id: 'last-order', score: 0.6, tokens: 100 },
];

pickMemories(memories, 500);

Try it in the editor

1// Rough estimate: about 4 characters per token.2const countTokens = (message) => Math.ceil(JSON.stringify(message.content).length / 4);34function trimContext(messages, maxTokens) {5  // TODO6  return messages;7}89module.exports = { trimContext, countTokens };

Legal notice and privacy Manage cookies