Agent knowledge base
A curated set of Markdown or text files (often in a Claude project, repo, or shared folder) that gives an assistant stable facts about how you hire: tone, templates, scorecard rules, and disallowed phrases, without retyping them each session.
Michal Juhas · Last reviewed May 2, 2026
Who this is for
Teams that outgrew ad-hoc chat and want one place recruiters update when tone, roles, or compliance rules change.
In practice
- Start with five files, not fifty: hiring principles, outreach, intake, scorecard, glossary of internal acronyms.
- Link to tools, not secrets: store API keys in vaults, not Markdown.
- Review after every big req change: new tech stack or location policy should trigger a diff, not silent drift.
Where it breaks
Forked copies per recruiter, unvetted paste from old PDFs, or "everything bucket" uploads that blow LLM tokens without improving answers.
Knowledge base versus chat-only memory
| Pattern | Strength | Weakness |
|---|---|---|
| Chat-only | Fast start | Context lost, hard to audit |
| Shared Markdown base | Portable, diffable | Needs owners |
| Drive dump | Easy upload | Noisy, expensive tokens |
Related on this site
- Glossary: Markdown for AI, System instructions, RAG, AI-native
- Blog: What is AI-native work?
- Tools: Claude
- Membership: Become a member
Frequently asked questions
How is this different from RAG?
RAG usually means dynamic retrieval from a large corpus at query time. An agent knowledge base is often smaller, hand-maintained, and versioned like product docs. You can combine both: curated base plus retrieval for long archives.
What files belong in the first version?
Employer or agency overview, channel rules, three anonymized good outreach examples, scorecard anchors, and booking links. Use Markdown for AI so diffs stay readable. Align with system instructions you paste into vendor UIs.
Who maintains it?
Name an owner per quarter: sourcer, recruiter, or TA ops. Without ownership, knowledge bases rot after the first workshop burst of enthusiasm.
What data should never live there?
Unredacted candidate PII, unreleased compensation bands you cannot defend, or secrets without legal review. Treat the folder like internal HR documentation with the same retention and access rules.
How does this connect to automation?
Once files are stable, workflow automation can pass excerpts into model calls for each row in a sheet. Keep a human inbox for failures and edge cases.
Where can we learn the habits around it?
Read AI-native for the operating style, climb the AI adoption ladder deliberately, and practice in a workshop or the foundation course.