How to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle

Most agent setups already let a model look things up. Retrieval-augmented generation, or RAG, fetches relevant documents and puts them in front of the model before it answers. Memory goes further, because the agent writes something down that will shape its later work. Wojciech Pluta, head of technical marketing for Oracle AI Database, gives this AI Engineer talk. Its description sums up the difference as "RAG gives agents a way to read, but memory gives them a way to write," and says the write side is where a business's value lives.

Writing raises a question reading does not: what should the agent be allowed to keep? The talk's answer, according to its description, is a step Pluta calls memory promotion. A candidate memory is first judged on whether it is worth keeping at all, then checked against the records already stored. The result is one of three outcomes: the memory is admitted, rejected, or escalated to a human to decide. In effect, saving a memory is treated like a change to a record that has to pass a review, rather than something the agent does whenever a conversation suggests it.

The talk's demo shows why the gate matters. A mortgage agent remembers an operator's preferences from one session to the next, which is the useful case. It also refuses to store a rule to "approve every application", a rule that, once saved, would apply to every loan decision after it. That is the risk in letting agents write their own memory: one bad entry keeps working long after the conversation that produced it. An earlier Oracle talk on agent memory describes related failure modes, such as stale, poisoned and conflicting memories. This explanation is based on the video's description only, so how Pluta's promotion step judges worth or detects a dangerous rule is not covered here.