The Cost of Memory Isn't What You Store
Storage was never the expensive part. The expensive part is a note that was true when you wrote it, isn't anymore — and that your agent still trusts with full confidence.
A Signal — one idea, checked against my own work.
We talk about agent memory like it’s a storage problem. Add a vector store, widen the context, keep more… But storage was never the expensive part. Storage is cheap and getting cheaper.
The expensive part is a note that was true when you wrote it and isn’t anymore — and that you still trust. A stale fact an agent believes with full confidence does more damage than a fact it never had, because now it’s acting on it. The survey literature on agent memory has gotten good at cataloguing where to put things — internal versus external, episodic versus semantic, agent- versus user-centric, a whole taxonomy of substrates. What the taxonomy can’t hand you is the judgment of when a stored thing has gone off.
That judgment is the actual work, and it’s a discipline, not a feature. Write the memory the same session the fact showed up, while the rationale is still attached to it. Keep the index slim enough to read, so you can see at a glance what you’re trusting. And check the world before you lean on an old note — treat a memory as a claim to verify, not a truth to assume.
None of that shows up on a storage bill. It shows up later, as an agent confidently doing the wrong thing because it remembered too well and re-checked too little. The cost of memory isn’t what you keep. It’s what you keep trusting past its expiry.
I wrote the file-based version of this discipline down — the setup I run under my own agents. It’s there for anyone who’d rather run it than rebuild it.
Sources: A Survey of Agent Memory in the Second Half — arXiv 2602.06052 · The Persistent Memory Bundle