# EP 21: Why Does My AI Miss Memories? Second-Pass Recall Explained - URL: https://www.agenticpod.net/episodes/ep21-second-pass-recall/ - Watch on YouTube: https://www.youtube.com/watch?v=ZKVfDvCMRsQ - Published: 2026-10-22 - Length: 5 min - Hosts: Stella (asks), Liam (answers). Both are AI hosts with AI-generated voices. ## Summary Your AI called memory and still gave a confident, wrong answer. The fix is one extra step: run the AI's own draft back through memory as a new query. ## Key takeaways - Retrieval is only as good as the query: a short prompt pulls back the obvious notes and misses the quiet one next to them. - Second-pass recall: draft the answer, send the draft back through get_context as a new query, revise if something new shows up. Cap it at two or three rounds. - It matters most before decisions, before retrying something that failed, in unattended scheduled runs and in coding agents before they change code. - Today it is an instruction to your AI; in n8n you can make it a workflow step so it is enforced, not hoped for. - Vilix AI returns recent messages and related past turns with timestamps, so the newest fact wins when two notes disagree. ## Chapters - 0:00 Intro - 0:40 The wrong answer - 1:28 The second pass - 2:35 Where it matters most - 4:12 The thing he forgot ## FAQ **What is second-pass recall?** You draft an answer, then run that draft back through memory as a new query. The draft has more detail than your short prompt, so it can pull back notes the first lookup missed. **How many passes should I run?** Two or three. Most of the value is in the first extra pass, and you stop when nothing new comes back. **Does it only work in chats?** No. It matters most in agents that act on their own, because a missed note becomes an action. In n8n it can be a loop step that checks whether anything new came back before the agent acts. **Which tools does it work with?** Any tool connected to the same Vilix AI memory over MCP: ChatGPT, Claude, Claude Code, Codex, Cursor and more. ## Resources - [Vilix AI, one shared memory for ChatGPT, Claude, Claude Code, Codex, Cursor, OpenClaw, Hermes and every MCP tool](https://vilix.ai) - [Connect Vilix AI in under two minutes per tool](https://vilix.ai/get-started) - [Technical guide with prompts](https://github.com/asly-97/ai-automation-workflows/tree/main/second-pass-recall) ## Transcript ### Intro Liam: Hey hey hey, welcome to Agentic Pod! I'm Liam. Stella: And I'm Stella! Liam, our founder asked his AI a simple customer question last Tuesday. It answered, confidently, and it was wrong. Why? Liam: Because the first memory lookup missed one note. And the fix took one extra step. Today: second-pass recall, what it is, and how to wire it up with Vilix AI. Stella: One extra step. I'm suspicious of anything that cheap. Liam: Good. Stay suspicious. ### The wrong answer Stella: Okay, so what actually happened? Liam: He asked Claude, connected to Vilix AI, a lookup question about his customers. The first answer sounded totally reasonable. It pointed at one person who looked like they'd signed up under a different email. Stella: And that was wrong? Liam: Yep. Plausible, but wrong. And here's the thing, the AI did call get_context. It did use memory. It just asked one question, with the user's short prompt, and got back what that one question matched. Stella: So memory worked, and the answer was still off. Liam: Right. Retrieval is only as good as the query. A short prompt pulls back the obvious notes. The quiet note sitting next to them never shows up. Stella: I feel like I do that to my own notes. So what did he do next? ### The second pass Liam: He said, use your own answer as the query. Take the draft, send it back through get_context, then answer me. Stella: Wait, the answer becomes the question? Liam: Pretty much. The draft has way more detail than the original prompt. Names, assumptions, a theory. Each of those can pull back a note the first query never touched. Stella: And it worked? Liam: It surfaced a note from two days earlier. An account count that had dropped. The first pass never returned it. Once the AI read that, the conclusion flipped, and the real answer came out. Stella: So the memory was there the whole time. Liam: It just needed the right trigger. And that's the idea. Pull context, draft your answer, then run the draft back through memory as a new query. If something new shows up, revise and go again. If nothing new shows up, it's safe to act. Stella: How many rounds? Forever? Liam: Cap it at two or three. Most of the value is in the first extra pass. Stella: Okay, but that sounds like a trick for chats. What about agents running on their own? ### Where it matters most Liam: That's where it matters most. Four places. Before a decision. Before you retry something that failed once. In scheduled, autonomous runs where nobody's watching. And in coding agents, before they change code. Stella: Because in those, a missed note becomes an action. Liam: Exactly. A chat gives a wrong answer and you notice. An agent does the wrong thing and you find out later. Stella: So how do I actually set it up? Liam: Today, it's an instruction. You tell your AI: after you draft your answer, call get_context again with the draft as the query, and repeat until nothing new appears. Vilix AI returns recent messages and related past turns with timestamps, so the newest fact wins when two notes disagree. Stella: And then save_turn at the end, so the next run knows what happened. Liam: Yep. get_context at the start, the second pass on the draft, save_turn at the end. And it works across ChatGPT, Claude, Claude Code, Codex and Cursor, because they all share one Vilix AI memory. Stella: What about the n8n people? The ones building agents, not chatting? Liam: Make it a workflow step. The agent drafts, a step calls get_context with the draft, a loop checks whether anything new came back, and only then does the agent act. Enforced, not hoped for. Stella: Not a prompt someone can forget. Nice. Liam: Now, one more thing from that day. ### The thing he forgot Stella: Okay, you can't say "one more thing" and stop. Liam: So while he was explaining this idea to his AI, it told him he'd already pitched this exact concept. On October ninth. In a voice call with a different AI, Maya. Stella: He forgot his own idea? Liam: Completely. Cross-tool memory remembered his own thought for him. Vilix AI had saved that call, and it came back in the second pass. Stella: Amazing, and a little embarrassing. Liam: Both. And he has a human way to explain it. Think of a song you played in your car every day on your commute. Hear it again years later, and the whole period comes back. One trigger, a whole web of memories. Stella: So agents should work like that. The draft is the song. Liam: That's it. The draft is the trigger. So here's your homework. Take your next important AI answer, ask it to run its own draft back through memory, and watch what changes. Vilix AI is linked in the description, with the setup guide. Have you caught your AI missing something? Tell us in the comments. See you next time!