Agent Memory
Make agents remember the right facts about a customer and a case for weeks, keep each case's memory sealed, and prove that memory helps rather than harms.
2 modules · 8 lessons~20 hoursUpdated 2026-10-30
- Lessons8
- Quizzes8
- Sources52
- Of study~20h
About this course
About this course
An advanced course on what agents should remember and how. You map working, episodic, semantic and procedural memory onto concrete mechanisms (the context window, compaction and context editing, CLAUDE.md and skills, the Claude memory tool and Managed Agents memory stores, and services such as Mem0, Letta and Zep), then read their benchmarks critically. The second module is about running memory in production: scoping stores per customer, case, organisation and agent under confidentiality duties and data-protection rules on retention and erasure (such as GDPR), write, update and forget policies with provenance, structured state shared by a fleet of per-step agents, memory poisoning, and evaluating whether memory pays for its retrieval cost. The running example is a 12-week case handled by per-step agents and a customer-facing assistant; every lesson has a worked example in Python, primary sources, a quiz and exercises you do with Claude.
About the running example
The examples follow a fictional property-law (conveyancing) firm in England and Wales: a 12-week leasehold flat purchase handled by per-step agents and a client-facing assistant. It's a case study, chosen because its documents are long, structured and high-stakes. You don't need a legal background, and the same patterns apply to contracts, insurance claims, invoices, support tickets or patient records. In the lessons, a matter is one client's case and a solicitor is the lawyer who signs work off.
What you'll learn.
- 1Classify any fact an agent needs as working, episodic, semantic or procedural memory and choose the mechanism that should hold it
- 2Configure compaction, context editing, CLAUDE.md, skills, the Claude memory tool and Managed Agents memory stores, and predict what each keeps and loses
- 3Compare Mem0, Letta, Zep and LangMem by what they actually do, and read LoCoMo and LongMemEval results critically
- 4Design memory scopes, retention and erasure that satisfy confidentiality duties and data-protection law such as GDPR, with provenance that makes deletion possible
- 5Implement write, update and forget policies, including conflict resolution when facts change and structured state shared by a fleet of agents
- 6Defend memory against poisoning and evaluate it with ablations that measure recall, staleness, leaks, latency and cost
2 modules, 8 lessons.
10h 5m of lessons
Module 1 · 4 lessons · 4h 50m
How agents remember
The kinds of memory an agent needs and the mechanisms that implement them: the context window and compaction, files, skills and memory tools, and dedicated memory services with their benchmarks.
Module 2 · 4 lessons · 5h 15m
Memory in production
Running memory in production: scopes, retention and erasure; policies for writing, updating and forgetting; shared state across an agent fleet and its security; and evaluating whether memory pays for itself.
A memory layer for a 12-week case
Build and evaluate the memory for a simulated 12-week case handled by per-step agents and a customer-facing assistant, with scoped stores, typed time-aware facts, shared state, poisoning defences, retention and erasure, and an ablation that proves it helps.
More in Documents, Retrieval and Memory.
- IntermediateDocument AI · Data pipelines
Document Ingestion Pipelines
Turn a case's PDFs, scans, forms and emails into verified, structured, privacy-safe data your agents and search can trust.
8 lessons ~16hFree - AdvancedRetrieval · RAG
Retrieval and Knowledge Systems
Build retrieval your agents can trust over a case's documents and a company's reference library, and measure that it finds what it should.
8 lessons ~24hFree