---
theme: seriph
title: "Lesson 10 — What Should an Agent Remember About a User?"
info: "English video course for AI Agents in Depth"
author: Bojie Li
transition: slide-left
mdc: true
lineNumbers: false
monaco: false
aspectRatio: 16/9
canvasWidth: 980
layout: cover
class: cover
---
Build · Chapter 3 · Memory and Knowledge
# What Should an Agent Remember About a User?
Memory levels, representations, evaluation, and privacy
Lesson 10 of 42 · 19 minutes · User Memory System; Three-Level Framework; Four Storage Formats; Privacy
---
Build · Chapter 3 · Memory and Knowledge
# Problems this chapter will solve
Lesson 10
What Should an Agent Remember About a User?
Lesson 11
Why Does Semantic Search Miss Exact Answers?
Lesson 12
Why Is One Retrieval Index Never Enough?
Lesson 13
When Should the Agent Decide What to Retrieve?
---
# Why this problem matters
Recall
Recover an explicit fact from a previous session.
Cross-session reasoning
Combine evidence from several interactions.
Proactive service
Notice a relevant need before the user repeats it.
---
# Three ideas to keep in view
Simple Notes
Atomic facts with little context
JSON Cards
Structured facts with evidence and scope
Advanced Cards
Conflicts, confidence, time, and applicability
---
# The book's visual model
Four strategies for representing user memory
---
# Store everything vs. Managed memory
Store everything
- High noise
- Privacy exposure
- Contradictory details
Managed memory
- Evidence-backed entries
- Conflict resolution
- Retention and sanitization
Memory is a governed knowledge system, not a transcript archive.
---
# A memory needs provenance
~~~json
{
"fact": "Prefers aisle seats",
"scope": "long-haul flights",
"evidence": ["session-18:turn-9"],
"confidence": 0.82,
"updated_at": "2026-08-03"
}
~~~
---
# Test the claim
3-1/3-23 min
Compare user-memory representations
Observe: What is extracted, how context is preserved, and how conflicts appear
3-32 min
Sanitize a memory-bearing log
Observe: Secrets removed while diagnostic structure remains
Demo budget: 5 minutes · one contiguous terminal block
---
class: course-terminal
---
Live demo
# Switching to the terminal
~~~bash
$ uv run python chapter3/user-memory/main.py --mode demo --memory-mode advanced_json_cards
$ uv run python chapter3/log-sanitization/main.py --demo
~~~
Run the command(s), narrate decisions, and point to the observation—not just the output.
---
# What the evidence supports
Finding 1
Memory quality must be evaluated at several capability levels.
Finding 2
Structured representations retain scope and provenance better than flat notes.
Finding 3
Privacy controls belong in the ingestion path, before persistence.
---
# Boundary → design rule
A detailed schema improves precision but increases extraction cost and schema-maintenance burden.
Store the minimum durable claim together with evidence, scope, confidence, and time.
---
# Continue the experiment
---
layout: center
class: text-center
---
Pause and apply
# Your turn
How should a memory system respond when a newer statement contradicts an older one?
---
layout: center
class: text-center
---
Next · Lesson 11
Retrieve external knowledge when exact words and semantic similarity disagree.
→
