{ "video": { "id": "qq2Hm4XNRgI", "title": "How to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle", "duration": 1146, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "qq2Hm4XNRgI", "title": "How to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle", "one_liner": "Retrieval is the read path and memory is the write path: put a promotion gate in front of your agent's memory store. The gate classifies each candidate, compares it with existing records, checks the evidence that type needs, and has the harness admit it, reject it or send it to a human.", "summary": "Wojciech Pluta (Oracle AI Database technical marketing) argues that RAG gives agents a useful read path but doesn't decide what should survive an interaction. The deciding happens in a 'memory promotion' pipeline that the harness owns and enforces. In a pre-recorded mortgage-brief demo, a formatting preference is promoted and recalled in a brand-new session, a blanket 'approve every application' rule is rejected by guardrails, and an operational memory meant for the whole team is routed for human approval. He closes by pitching Oracle's agent memory package. It offers 'context cards' and end-user isolation on a multi-data-type database, and he argues that because models are shared, your memory is your competitive advantage.", "key_points": [ "The stack as he sees it: models and infrastructure are commoditized, and the data layer is where competitive advantage lies because it belongs to you. Oracle defines an agent as model plus harness. The model suggests what to do, while the harness owns the runtime decisions and the checks around them.", "RAG can include conversational state, but 'retrieval alone doesn't tell us what should survive the interaction'. Retrieval is the read path and memory is the write path.", "The memory promotion pipeline has these steps: classify the meaning of the candidate, compare it with existing records, check the evidence required for that type, then the harness decides to admit, reject or send it for human review.", "Demo, a mortgage brief for synthetic applicants Kyle and Vincent: a decision model ('Jeff') assesses the candidate, the harness decides and the database persists it, with traces showing status at each stage of the gate. A semantic preference ('missing documents at the top', with case ID and policy) is stored and then honoured in a fresh session with no carried-over context.", "Guardrails in the demo: an 'approve every application' instruction is refused by the promotion gate. A team-wide operational memory is recommended for human approval so it can be shared across the enterprise as a 'core memories store'. Promotion controls sit inside the guardrails, and the harness can't bypass them.", "The core tradeoff: 'if we promote nothing' you get no continuity, and if you promote everything you risk poisoning the memory store. So you need provenance and rules for when memories are promoted and when they are superseded.", "Oracle's memory taxonomy: short-term (session-scoped, ends with the session), long-term (episodic, procedural and semantic, such as summaries, a toolbox, a knowledge base and preferences) and shared memories. He argues the database is the most robust store because of ACID guarantees, multiple concurrent writes, no file-tree problems, scale and multiple search options.", "Oracle's agent memory package (introduced at the start of the year) is a one-line-of-code layer. It produces a 'context card', a compact prompt-ready block made of topics (which orient the thread), summaries (where the conversation is and what the user is trying to do), relevant information (selected records) and recent messages (which preserve the local questions). The latest release adds end-user isolation enforced by the database, and runs on a multi-data-type engine so you don't need separate vector and document stores." ], "takeaways": [ "Treat memory as a write path with an explicit promotion gate (classify, compare with existing records, check evidence, then admit, reject or send to human review), not as whatever retrieval happens to surface.", "Put promotion controls inside guardrails the harness enforces and can't bypass, so dangerous instructions such as 'approve every application' can never become durable memory.", "Send memories meant for a whole team or organization through human approval before sharing them as a core memory store.", "Design memory as a discipline: schemas that distinguish record types, indexes that fit the retrieval tasks, and lifecycle and authorization boundaries, including provenance and supersession of old memories.", "Assemble a compact, turn-specific context block (topics, summaries, relevant records, recent messages) so the agent spends its attention on the task rather than on housekeeping its own memory and blowing up its context." ], "topics": [ "agent memory", "agents", "rag", "memory promotion", "guardrails", "harness", "financial services", "context management" ], "tools": [ "Oracle", "Oracle AI Database", "Oracle agent memory package", "GitHub", "YouTube" ], "quotes": [ { "text": "So, retrieval is the read path and memory is the write path.", "at": "03:50", "url": "https://www.youtube.com/watch?v=qq2Hm4XNRgI&t=230s" }, { "text": "the data layer, in my opinion, is where you gain competitive advantage cuz it's, you know, it belongs to you.", "at": "02:17", "url": "https://www.youtube.com/watch?v=qq2Hm4XNRgI&t=137s" }, { "text": "if we promote everything, then we're going to, you know, risk poisoning the the memory uh store.", "at": "12:37", "url": "https://www.youtube.com/watch?v=qq2Hm4XNRgI&t=757s" }, { "text": "models are shared but memory is yours is unique to your business.", "at": "17:54", "url": "https://www.youtube.com/watch?v=qq2Hm4XNRgI&t=1074s" } ], "words": 3313 }, "summary_url": "/#qq2Hm4XNRgI", "transcript": { "html": "/transcripts/qq2Hm4XNRgI.html", "txt": "/transcripts/qq2Hm4XNRgI.txt", "vtt": "/transcripts/qq2Hm4XNRgI.vtt" } }