Wojciech Pluta, head of technical marketing for Oracle AI Database, presents a 19-minute AI Engineer talk on choosing and governing persistent agent memory.

Decide what earns a place in memory

  • Retrieval-augmented generation supplies relevant information to a model; Pluta frames memory as deciding what new information to retain for later.
  • A promotion pipeline classifies a candidate, compares it with existing records, checks required evidence, and admits, rejects, or sends it to a human.
  • Saving nothing loses continuity; saving everything risks stale, conflicting, or poisoned memory.
  • Provenance, supersession, access controls, and lifecycle rules matter alongside retrieval.

A mortgage-agent demonstration

  • In a prerecorded sandbox, the operator requests a consistent brief format, with missing documents at the top.
  • A new session retrieves that preference and applies it to another synthetic application.
  • A request to approve every application is rejected by the promotion gate; broader shared memory can be sent for human approval.
  • This is an illustrative demo of configured controls, not evidence that all unsafe instructions are reliably blocked.

Keep context compact

  • Pluta distinguishes short-term session state, long-term records such as preferences and procedures, and shared organizational memory.
  • A context card assembles relevant records, summaries, topics, and recent messages for the current turn.
  • Oracle advocates database storage for transactions, search and access controls; its agent-memory package is the product he presents, not the only possible architecture.

“Retrieval is the read path and memory is the write path.” — Wojciech Pluta, 3:50