Building Agents with Memory

Agents need memory to be truly autonomous. Without it, they are just loops.

The Loop

A memory-enabled agent loop looks like this:

1. Observe: Receive input from environment.
2. Recall: Query OpenMemory for relevant past experiences.
3. Plan: Use LLM to decide action, conditioned on memory.
4. Act: Execute tool or API call.
5. Reflect: Store the result and the reasoning back into OpenMemory.

Code Snippet

async function agentLoop(input: string) {
  const context = await mem.query(input);
  const plan = await llm.plan(input, context);
  const result = await tools.execute(plan);
  await mem.add(`Action: ${plan}, Result: ${result}`);
}