An agent that forgets everything wastes time. It rereads the same sources, repeats the same bad idea, and spends the same API money twice. That is why Letta caught my attention.
I did not want a magic memory blob, though. If an AI can casually rewrite what it remembers, it can turn a guess into a fact. That is dangerous when the subject is money.
Memory points to evidence
In this lab, Letta holds small typed pointers: the job an agent was working on, the evidence record it used, the strategy version it handed off, and the result it needs to revisit. The real source, test, decision, and trade episode live in PostgreSQL with timestamps and hashes.
If memory and the database disagree, the database wins. If the pointer is stale or missing, the agent stops and asks for the real record. Memory cannot change risk limits, authorize orders, or turn an uncertain strategy into a winner.
That still makes it useful
The Librarian can avoid rediscovering the same paper. The skeptic can remember why a strategy failed. The Chief of Staff can resume a handoff after a restart. We get continuity without pretending an AI recollection is proof.
Remember where the evidence is. Do not remember a prettier version of the evidence.