How we taught agents to remember without forgetting what matters

Most AI assistants forget everything the moment a conversation ends. The obvious fix is to store everything. We tried that first, and it made our agents worse.

When an agent can recall every message, every document and every decision, it has to choose what is relevant for the task at hand. Retrieval systems tuned for search are good at finding text that looks similar. They are much worse at finding text that matters.

Recall is a ranking problem

We reframed memory as a ranking task with three signals: semantic similarity, recency, and a learned importance score that rises each time a memory changes an outcome. A pricing decision referenced in five later conversations outranks a one-off brainstorm, even if the brainstorm is more similar to the query.

score = 0.55 · similarity
      + 0.20 · recency(decay=30d)
      + 0.25 · importance

score = 0.55 · similarity
      + 0.20 · recency(decay=30d)
      + 0.25 · importance

score = 0.55 · similarity
      + 0.20 · recency(decay=30d)
      + 0.25 · importance

What changed

Across 1,200 internal evaluation tasks, agents using ranked memory completed 23% more tasks correctly and asked 41% fewer clarifying questions. Teams in the beta noticed the difference within a week.

“It stopped asking me things I’d already told it. That was the moment it started feeling like a colleague.”

What’s next

We are working on shared memory between agents, so a finance agent can learn from a decision your sales agent saw. Expect an early preview before the end of the year.