←Articles

The agent that cites itself

An agent can end up trusting its own old guess, because it reads that guess back in the same voice as your facts.

On Tuesday your agent makes a guess while it works. It is a good guess, and it saves it.

A week later it recalls that line. It reads it in the same voice as the facts you gave it, so it treats the guess as one of them. It builds on it and saves the new conclusion too.

Do that a few times and the first guess is now the root of a whole family tree. Every note in it points back to something the agent made up on a Tuesday, and nothing in memory says so.

Nobody attacked anything. The agent just kept quoting itself, one reasonable step at a time.

The fix is boring and it works: a memory should know where it came from. In SmartMemory every memory carries its origin, so something you said is not the same kind of thing as something the agent guessed. The guesses can still turn up when you search for them, but they do not get recalled with the weight of your own words. And for a decision, you can ask for its full chain of evidence and see exactly what it stands on.

Think of it as footnotes for your agent's own thinking.

Track where your agent's memories come from: pip install smartmemory, then the quickstart at https://docs.smartmemory.ai/smartmemory/getting-started/quickstart