Search and memory

Ask Dex "what did we decide about pricing in the spring?" and it can usually just tell you. This page explains the machinery behind that — how a fundamentally forgetful AI ends up with a better memory of your work than you have — and how to upgrade its search from keywords to meaning.

The memory model, honestly

The AI underneath Dex remembers nothing between sessions. Dex's memory is entirely files plus discipline:

• Your vault is the long-term memory. Every meeting, decision, task, and person page is a file. If it's written down, Dex can find it; if it was only ever said aloud, it's gone. That's why capture is the habit Dex protects most.
• Session context is the working memory. When you start a conversation, hooks load your priorities and recent patterns; when you mention a person or company, their context is injected automatically. Dex "remembers" you by rereading the right files at the right time.
• Learnings are the self-improving layer. Corrections, preferences, and mistake patterns captured during reviews get written down and resurface in future sessions — so a mistake made in March can shape behavior in July.

The practical takeaway: tell Dex things. Every fact you mention becomes permanent, searchable memory.

Ordinary search: solid, literal

Out of the box, Dex searches by keywords and file structure — exact words, filenames, folders. For known terms ("the Northwind contract") it's fast and reliable. Its blind spot is vocabulary: search "customer retention" and it won't find the note that said "churn."

Semantic search: finds what you meant

/enable-semantic-search upgrades Dex to search by meaning. After it's on, "customer retention" finds notes about churn, cancellations, and NPS; "product-led growth" finds the note that said "self-serve motion." Skills across Dex quietly get better too — task matching, meeting linking, person lookup, and finding related projects all use it once available.

How it works, in one paragraph: your notes are converted into mathematical representations of meaning (embeddings), your question gets the same treatment, and results are ranked by closeness of meaning — then re-ranked by a second model for true relevance. Three small AI models do this work.

What matters about the setup:

• Everything runs on your machine. The models are downloaded once (about 2GB total) and your notes never leave your computer — no cloud, no API key, no account.
• It's a guided, consented install. Dex checks your disk space, explains exactly what it's downloading, and asks before touching anything.
• It organizes itself around your vault. Rather than one giant index, Dex scans your vault and proposes purpose-built collections — people, meetings, projects, goals — which makes results sharper.

If you skip it, nothing breaks: every feature that prefers semantic search falls back to keyword search automatically.

Keeping the index fresh

The index updates as you work, but if results ever feel stale (files you added recently not showing up), the fix is one line — ask Dex to refresh the search index, or run:

bashqmd update && qmd embed

Re-running /enable-semantic-search on an existing setup acts as a health check: it reports stale collections and suggests exactly that refresh.

Getting good answers

• Ask questions, not keywords. "What did Sarah think about the timeline?" beats "Sarah timeline."
• Ask across time. "Have we discussed this before?" is one of the highest-value questions in Dex — it's where the compounding memory pays off.
• Check the source. Answers come from your files, and Dex can show you which — everything is inspectable.

The "ask across time" move deserves its own prompt — use it before any significant decision:

Before I commit to [DECISION]: search everything. Have we faced something
like this before? What did I decide then, what was my reasoning, and what
actually happened afterward? If past-me and present-me disagree, lay out
both cases and tell me what changed — the situation, or just my mood.

Make it your own: this is the prompt that makes six months of captured meetings suddenly pay rent. The longer you've used Dex, the sharper it gets.

Watch it work:

First, fetch and read the page I'm referring to: https://heydex.ai/help/search-memory.md

I just read the Dex Guide's page on search and memory. Prove it to me:
tell me three true, specific, useful things from my own vault that I have
probably forgotten — a commitment, a decision, a signal — each with the
source file and date. No generic observations; surprise me. Then, based
on what's actually in my vault, give me the one question I should ask
you every week that I'm not asking.