Your agent works in the demo, then forgets you tomorrow. Models are stateless by design, and memory is the harness you build around them.

One question picks the store: do you know the key, only the meaning, or the relationship? That is the difference between key value, vector, and graph memory. Live demos show each one winning and breaking, so you leave knowing which memory your agent needs.

What you'll learn:
• Pick key value, vector, or graph with one test: do you know the key, the meaning, or the relationship
• Build the write path and read path that move facts from a conversation into a store and back into the window
• Decide what to keep with an agent-owned prompt, typed stores, or a managed extractor, and know who holds the policy in each
• Block poisoned writes with a gate and remove what got in, sized to the blast radius of your store
• Record decision traces so a wrong source can be traced to every decision that trusted it

Outline:
• The failure: works in the demo, forgets you tomorrow
• The one test that picks the store
• Key value: name the key
• Vector: know only the meaning
• Graph: follow the relationship
• The layer on top of the store
• Choosing, and the honest limits