YouTube Script: Pattern 3 — Memory

Title: "Your AI Agent Forgets Everything Between Sessions — Fix It In 10 Seconds" Duration: ~7 minutes Style: Screen recording + talking head


HOOK (0:00 - 0:40)

[SCREEN: Repeating clip — same agent question across 3 different dates: "What Python version should I use?"]

VILIUS (voiceover): "Session one: what Python version? Session two: what Python version? Session three: what Python version? This isn't the agent being stupid. It's the agent having no memory. And the fix takes ten seconds."

[SCREEN: Cut to talking head]

VILIUS (on camera): "I tracked this. Over three weeks, I spent thirty minutes just correcting basic facts the agent should have known. Python version, test command, project structure. Every session, starting from zero. Here's how to make it stop."


THE PROBLEM (0:40 - 1:30)

[SCREEN: Animation — agent sessions as blank slates, each one being filled in from scratch]

VILIUS (voiceover): "Most people treat every agent session as a fresh start. They don't save anything. The agent learns, does the work, and then forgets everything when the session ends. Next session — blank slate. You re-explain. You re-correct. You waste time."

[SCREEN: Sticky note on monitor: "REMEMBER TO TELL AGENT: python3.11, tests/ dir, don't commit to main"]

VILIUS (on camera): "By week four, I had a physical sticky note on my monitor with six things I had to tell the agent every session. That sticky note is memory. But it's in the wrong place. It should be in the agent."


THE PROTOCOL (1:30 - 3:30)

[SCREEN: Terminal — agent makes mistake, user corrects it]

VILIUS (voiceover): "Here's the correction protocol. It takes ten seconds and it transforms how your agent works over time. Watch."

[SCREEN: Step-by-step]

VILIUS (voiceover): "Agent does something wrong. You correct it. Then — and this is the key — you say 'Remember that.' The agent saves the fact to persistent memory. Next session, it reads that memory. It never asks again."

[SCREEN: Show memory being saved — text entry appearing in a memory file]

VILIUS (on camera): "The most valuable memory entry is a correction. Every time you correct your agent on something, save it. Python version, test command, communication style, project conventions. Ten seconds now saves two minutes every session forever."


WHAT TO SAVE (3:30 - 5:00)

[SCREEN: Three-column graphic — Corrections Preferences Environment]

VILIUS (voiceover): "Three things go into memory. Corrections — every time you fix the agent. Preferences — communication style, autonomy level, tool choices. Environment facts — versions, paths, dependencies. That's it."

[SCREEN: Red X over "Task progress," "Session outcomes," "Raw data"]

VILIUS (voiceover): "What doesn't go in memory: task progress, session summaries, raw data dumps. Those belong in your project tracker. Memory is for facts that will still matter in three months."

[SCREEN: Memory vs Skills comparison]

VILIUS (on camera): "One more distinction. Memory says that something exists. 'We use pytest.' A skill says how to use it. 'Run tests with pytest -n 4 -v, check coverage with --cov-report.' If you're saving procedures to memory, stop. That's Pattern 2 — skills."


THE NUMBERS (5:00 - 6:00)

[SCREEN: Before/after graphic]

VILIUS (voiceover): "Before memory: six corrections per session on average. Python path, test command, preferred tools, communication style, project structure, git workflow. About two minutes per session explaining the same things."

[SCREEN: After memory — zero corrections]

VILIUS (on camera): "After memory: the agent reads it all at the start of every session. Python version — known. Test command — known. Communication style — known. Correction rate dropped from six per session to half a correction. Over five months, that's hundreds of corrections that never happened."


OUTRO (6:00 - 7:00)

[SCREEN: Works With Agents logo]

VILIUS (voiceover): "Memory gives your agent durable context. It knows your environment, your preferences, your corrections. In the next module, we add Decision Protocols — rules for when the agent acts without asking. This is where autonomy begins."

[SCREEN: Call to action — "Memory checklist at workswithagents.com/learn"]

VILIUS (on camera): "Your homework: audit your last three sessions. What did you correct the agent on? Save those five things to memory right now. Start a fresh session tomorrow and count the corrections. I bet it's not zero yet — but it'll be close."


PRODUCTION NOTES

  • B-roll: Terminal showing correction protocol in action, memory entries being saved
  • Graphics: "What to save" three-column graphic, memory vs skills comparison card
  • Music: Minimal, ambient
  • Captions: Standard

← Back to all patterns

Spotted something?

Suggest an improvement, report an error, or just say hi.