Build a personal AI knowledge base from your own files

Step 1: Choose the corpus, not the shiny app

Pick sources that already exist:

• Meeting notes and decision logs you own
• Project READMEs and design docs
• Personal SOPs and checklists
• Export of notes from one notebook tool

Exclude for v1:

• Downloads dumps
• Duplicate “final” PDFs
• Anything with secrets or customer PII you should not store in that tool
• Huge archives you have never opened

A tight 50–200 pages of truth beats 10,000 junk files.

Step 2: Clean names and dates

Before any model feature:

• Rename notes.pdf to pricing-decision-2026-09-12.md
• Put a one-line title at the top of each doc
• Delete known-stale copies
• Keep one “current” folder and one “archive”

Retrieval surfaces noise as eagerly as signal. Filenames are half the battle.

Step 3: Pick a simple access pattern

For personal use, three patterns are enough:

1. Paste on demand — open the file, put the relevant section into the user message
2. Project / folder chat — tools that let you attach a workspace of files
3. Light retrieval — product features that search your files before answering

Start with (1) or (2). Graduate to (3) when context windows get annoying. Do not buy enterprise search to organize your own Notion export.

Step 4: Write a personal ask template

Reuse:

Use only the attached / selected files.
Question: […]
If the answer is not in the files, say “Not in my notes” and list what is missing.
Cite file names in the answer.

That “Not in my notes” line is the difference between a grounded knowledge base and unconstrained generation.

Step 5: Weekly hygiene (20 minutes)

Every week:

• Add new decision notes from meetings
• Archive one outdated doc
• Spot-check one answer: did it cite the right file?
• Remove any file that should never have been there (secrets, customer exports)

Without hygiene, your “brain” becomes a junk drawer with a search box.

Privacy and device rules

• Prefer local or business accounts you understand for work-adjacent notes
• Strip customer identifiers before files enter a personal base
• Do not mix highly regulated records into a hobby vault
• Know how to delete a file from the tool, not only from disk

If the file would be awkward on a lost laptop, it is awkward in a synced model-backed vault too.

What good answers look like

A useful reply:

• Points to a file name and section
• Separates “in my notes” from “general advice”
• Admits gaps

A useless reply:

• Sounds sure with no citation
• Mixes three projects into one answer
• Invents a policy you never wrote

When you see useless, fix the corpus or the template—not your trust in vibes.

Worked example (pattern only)

You keep decision-logs/, sop/, and project-briefs/. You ask: “What did we decide about the pricing page owner?” The tool cites pricing-decision-2026-09-12.md and names Alex. When you ask about a vendor you never documented, it says “Not in my notes.” That second answer is success.

Common mistakes

• Dumping your entire Downloads folder on day one
• Asking product questions against personal diaries
• Never deleting superseded SOPs
• Expecting perfect recall without citations

Pair with paste vs retrieval

If most questions point at one known file, keep pasting. If you keep asking “where did we write that?”, graduate the same curated folders into folder chat or light retrieval—without dumping Downloads into the index. The knowledge base is the corpus discipline; the feature is secondary.

Short close

Curate a small corpus, demand citations, admit gaps weekly—that is a personal AI knowledge base that stays useful.