MICROSOFT TRIED TO BUY IT. HARVARD AND THREE US UNIVERSITIES RELEASED IT FOR FREE a team from Harvard, Michigan State, Minnesota, and Georgia ran the first hard study on how agent memory behaves over time. the finding is uncomfortab they call it the "experience-following property": when a new task... looks like a memory the agent pulls up, its answer copies that memory's answer almost 1:1 (correlation up to 0.95). the agent isn't reasoning - it's imitating that breaks in two ways: > error propagation - one wrong memory repeats on every similar task after it > misaligned replay - "successful" past runs turn out low-value or misleading when reused so "just add more memory" quietly makes agents worse. what fixed it: > store only high-quality experiences, not everything > use future task outcomes as free labels to delete low-value memories > pair selective adding with utility-based deletion the payoff: ~10% absolute accuracy gains, a cleaner memory bank, and agents stable under distribution shift no company owns this. they dropped the whole thing open-source, for free bookmark this. the next edge in AI agents isn't more memory - it's cleaner memoryshow more

0.95 correlation shows agents copy past errors instead of reasoning

What you really want to do though is to select the memories to retain and then bake them into the model weights. For the procedural types of memories, we constructed data that did the task correctly and then tuned that.

0.95 correlation is basically copying your own homework

An agent is no different from a Human 🤯. Make it make sense. The minute you begin to talk about agents you're no longer talking about AI. Hey AI don't need agents.

This is why ‘save everything’ is a dangerous default. An agent needs provenance, confidence, and a way to reconcile conflicting information; otherwise it just turns old mistakes into repeatable behavior.

Cleaner memory over more memory, that's a smart finding.

Editable memory is especially important here, giving agents a way to correct or delete memories when they prove unreliable.

Turns out more memory can actually make agents worse 👀

Unsettling but valuable evidence on agent memory.

i`m note even surprised

“Just add more memory” quietly making agents worse is the counterintuitive finding here. Selective storage plus utility-based deletion for a 10% accuracy jump, free, is the real headline.

This is why a retrieved memory should be treated as a hypothesis, not an instruction. If it can’t lose authority after a bad counterfactual, “more memory” just means stronger path dependence.

let these Harvard guys try Claude + Obsidian they'll be blown away

The uncomfortable part is what happens when an agent’s old experience conflicts with current policy or source state. Do you model expiry and contradiction explicitly, or just rely on retrieval ranking?

More memory can literally make the agent worse.