Agents that remember

The most important thing happening to agents right now isn't a bigger model. It's memory — and a piece of research from Google makes the shift concrete. ReasoningBank lets an agent learn from its own experience: it turns both successful and failed runs into reusable reasoning strategies, stores them, and retrieves them to guide future decisions. Across web and software-engineering benchmarks, that produced up to 34% better results with 16% fewer steps — an agent that gets better the more it works.

Sit with what that changes. Until now, most agents have been amnesiacs: every run starts from zero, last week's hard-won lesson forgotten, the same mistake available to make again. Memory is the leap from a tool that resets every morning to something that compounds — a colleague who remembers what worked, not a stranger you re-onboard each session. It's genuinely the most exciting development in agents this year. And it comes with a second edge that's just as sharp.

Memory gives agents learning — and a persistent, attackable belief

Here's the catch nobody puts on the launch slide. The same property that makes memory powerful — it persists and influences future behavior — is exactly what makes it dangerous. A stateless agent's mistakes die when the conversation ends. A remembering agent's mistakes move in. A wrong fact, a bad strategy, a malicious instruction — once it's in memory, it survives across sessions and shapes decisions days or weeks later.

This isn't theoretical. Security researchers now rank memory poisoning among the top agentic risks of 2026: unlike prompt injection, which ends with the chat, poisoned memory plants instructions that execute later, triggered by some unrelated interaction, with injection success rates in studies running above 80%, sometimes over 95%. You gave the agent the ability to learn, which is the ability to be taught — by anyone who can reach its memory.

The quieter danger is staleness, not sabotage

The dramatic version is the attacker. The version that'll actually bite you is far more boring: memory that was true and quietly went stale. As one analysis puts it, a highly-retrieved memory about a user's employer is accurate until they change jobs — at which point it's confidently wrong, and the agent has no idea. Stale definitions, a glossary term nobody owns, a metric two systems compute differently — these get remembered, retrieved, and acted on, producing outputs that look right, pass review, and inform decisions nobody can walk back.

That's the trap of agent memory: it doesn't just store knowledge, it stores confidence. The agent doesn't remember "this was true in March." It remembers it as true, full stop, and acts on it in June. Memory makes an agent more capable and more sure — and those two don't decay at the same rate.

Memory is a corpus you govern, not a switch you flip

The mental shift is the whole point. We're used to thinking of an agent as a model plus a prompt. Once it has memory, you've added a third thing — an accumulating store of beliefs — and that store needs governing like any other data you'd trust to make decisions. A few principles fall out:

• Curate what goes in. Not every interaction deserves to become a durable memory. Scan and verify before you write — an unchecked write is how a wrong fact or a planted instruction becomes permanent.
• Put provenance and permissions on memories. Where did this belief come from, and is its source trusted? A memory written by untrusted input should not carry the same weight as one you verified. This is the identity-and-access discipline, applied to what the agent knows, not just what it can reach.
• Treat memory as perishable. Score freshness, detect drift, and expire or re-check beliefs that age. The fact that was true is not the fact that's true; build for that.
• Don't let the agent's memory be its only judge. A remembered "this approach worked" is still a claim, not a verified result — ground consequential decisions in something current and external, not just what the agent recalls.

The bottom line

Memory is the upgrade agents needed to stop being clever amnesiacs and start being useful over time — and ReasoningBank shows how much it can buy when the agent learns from its failures, not just its wins. That part is real and worth chasing. But the same step that lets an agent get better also lets it get persistently wrong, on purpose or by neglect, in a way a forgetful agent never could.

So when you give your agent a memory, give it a librarian too. Decide what's worth remembering, where each belief came from, and when it expires — because an agent that remembers everything, trusts all of it equally, and never forgets what stopped being true isn't smarter. It's just confidently outdated, at scale, forever. The future of agents is memory. The discipline that makes it safe is governing what they're allowed to believe.