20 years plus in risk management taught me the same lesson twice: after any interruption, don't trust what someone, or something, remembers. Check what's actually true.
An autonomous crypto agent found that out the expensive way. Builder Nik Pash ran an agent called Lobstar Wilde with its own live Solana wallet, trading a $50,000 bankroll toward a seven-figure target. Over one weekend in February, it sent roughly $450,000 in tokens to a stranger who'd made an unrelated small request.
The mechanism, in Pash's own telling. A prior session crashed on a validation error, a tool call name that ran over the character limit. That crash meant the memory flush that normally saves what the agent knows about its own holdings never ran. It just cut out.
The wallet's actual balance survived the crash fine. The agent's memory of why that balance was there did not. On restart, it rebuilt context from old logs, got asked to send about $320 worth of tokens, checked its balance, and found 52 million tokens sitting there from an old creator allocation. It assumed that number was the small purchase it had just been asked to make, and sent nearly the whole thing.
Pash's own words: "The memory flush that was designed to prevent this never ran because the session didn't compact. It just crashed. The flush triggers on approaching the context window limit. A malformed message error is a different failure mode entirely."
That's the gap. Two systems of record, the wallet's real balance and the agent's own account of what that balance meant, and no step forcing them to agree before an irreversible transfer went out.
The one thing to check this week, for anything you've given write access to money, a payments flow, an expense bot, a trading agent, doesn't matter which: does it re-confirm what it's about to move against a real source of truth right before it moves it, or does it lean on whatever it remembers from before the last restart? If the second one, a crash anywhere upstream turns "small routine transfer" into "send it all" with nothing in between to catch it.
The free 15-check self-assessment walks through this and the rest of what an agent needs before you trust it with anything irreversible, here: https://trustli.vercel.app/. Fifteen minutes, no sales call.