Episode 9: Skill 3 — Knowledge Agent
This episode's scenario: The blog project has been in development for 3 months with 500+ Observations. You want to ask Claude about your authentication system details, but search results are scattered across different sessions. You need a way to package all "auth" memories into a focused knowledge vault.
9.1 Why Knowledge Agent?
As projects grow, Memory Search hits its limits:
| Project Stage | Observations | Search Quality |
|---|---|---|
| Week 1 | ~50 | ✅ Precise |
| Month 1 | ~200 | 🟡 OK, occasional noise |
| Month 3 | ~500+ | ❌ Results too scattered for a complete picture |
Knowledge Agent's solution: Pre-filter and organize related Observations into a Corpus (knowledge vault), then query that vault directly.
9.2 Three Core Operations
build_corpus — Build a Knowledge Vault
build_corpus(
corpus_name: "auth_system",
query_tags: ["auth", "JWT", "login", "OAuth"]
)
Filters 500+ Observations down to ~35 auth-related ones, organized into a coherent knowledge body.
prime_corpus — Activate a Vault
prime_corpus(corpus_name: "auth_system")
Loads the corpus into Claude's current context. Now any auth-related questions get precise, history-based answers.
query_corpus — Ask the Vault
query_corpus(
corpus_name: "auth_system",
question: "What's our JWT token TTL? How does refresh work?"
)
Returns an answer grounded in actual Observations, not generic Claude knowledge.
9.3 When to Use search vs Knowledge Agent?
| Scenario | Recommended | Why |
|---|---|---|
| Find a specific bug fix | search |
Clear target, quick lookup |
| Understand a module's overall design | Knowledge Agent |
Needs synthesis across many records |
| Answer "why did we design it this way" | Knowledge Agent |
Needs causal chain and full context |
| Quick detail lookup | search + get_observations |
Precision targeting |
| Explain a subsystem to a new teammate | Knowledge Agent |
Needs coherent narrative |
9.4 Corpus Naming Convention
| ✅ Correct | ❌ Wrong |
|---|---|
auth_system |
Auth System (no caps/spaces) |
database_design |
Database Design (no caps) |
blog_frontend |
blog-frontend (use underscores) |
Rule: lowercase + underscores (snake_case).
Hands-On Exercise
- Ensure the blog project has 5+ sessions
- Build your first corpus:
"Build a knowledge vault about database design with tags: Prisma, schema, database, model" - Activate and query it:
"What tables does our database have? What are their relationships?"
Coming Up Next
With memory search, code exploration, and knowledge vaults, Claude can "remember the past" and "read code" effectively. Next: how to make it "plan the future" — Make Plan + Do.