Agent memory is mostly hoarding

Strong take: most of what gets sold as "agent memory" is a database with good marketing.

The pitch is seductive. Give the agent a memory store, let it write everything it sees, and it will get smarter over time. In practice what you usually build is a junk drawer that grows forever. Stuffing every past message and tool result into context is not memory. It is hoarding.

Real memory is knowing what to forget!

Accumulation is not recall

Human memory is impressive less for what it keeps and more for what it quietly drops. You do not remember every sentence of every conversation, and that is a feature. The signal survives because the noise decays.

Most agent "memory" does the opposite. It accumulates. Every run appends more, relevance drops, retrieval gets noisier, and the cost per call creeps up while quality creeps down. The agent is not remembering more. It is carrying more.

What real memory needs

The hard part is not writing to a store. Any database does that. The hard part is the editorial layer on top: deciding what is worth keeping, what should be summarized, and what should be dropped on purpose.

• Selective retention. Keep the durable facts, decisions, and preferences. Let transient chatter expire.
• Compression over accumulation. Ten messages that resolved into one decision should be stored as the decision, not the ten messages.
• Active forgetting. Stale state should be removed deliberately, not just buried under newer state. An agent that cannot drop yesterday's wrong assumption will keep acting on it.

// Memory is an editorial decision, not a write call.
function remember(event: Event, store: Memory) {
  if (isDurable(event)) store.upsert(distill(event)) // keep the essence
  if (supersedes(event, store)) store.forget(event.topic) // drop the stale
  // everything else is allowed to decay
}

Why this is the interesting problem

Selective retention and deliberate forgetting are genuinely open problems, which is exactly why "remember everything" is such a popular shortcut. It demos well and ships fast. It just does not scale past the demo, because relevance and cost both move the wrong way as the store grows.

So when someone shows you agent memory, ask the uncomfortable question: can it forget? Until the answer is yes, it does not remember. It accumulates.