Run your first query

Read this if you’ve finished setup and want to see OMem actually answer something. By the end you’ll have ingested some of your work and pulled your first result back out.

Step 1 — Ingest

Run one pass over the kinds you enabled:

omem ingest

This walks each item through the pipeline — discover it, parse it to Markdown, describe any images, curate a clean wiki page, index it. The first run does the most work (it’s seeing everything for the first time); later runs are quick because OMem only touches what changed.

1Read any format

parse

2Turn pictures into words

vlm + ocr

3Tidy it into one clean page

curate + save

4Make it findable

index

Your work lives in formats that don’t become clean text on their own — and most tools quietly drop the hard parts.

BEFORE

a .pptx: slides, speaker notes, an embedded chart, a merged-cell table

→

AFTER

clean Markdown — notes kept, the chart extracted, the table intact

A dedicated parser per format (PowerPoint, Excel, Word, PDF, scans, …) turns the file into Markdown reliably. It deliberately keeps what generic converters lose: embedded images, merged cells, speaker notes. This is deterministic — the same file produces the same Markdown years later.

Don't reach for --now

There’s a flag, --now, that forces a full re-scan. You don’t need it here — a plain omem ingest already ingests everything in scope on the first run. On a large mailbox --now can trigger a lot of LLM calls and real cost; the lifecycle page explains why. For now, just omem ingest.

Step 2 — Query

Ask a question in plain language:

omem query "Q3 budget review"

You get a ranked list of hits — each with a relevance score, the kind of work it came from, a short page-id, and a one-line summary:

↻ replay

$

Step 3 — Read the result

Each hit is one line — [score] kind page-id title — with its one-line summary indented underneath:

• score (in brackets) — how relevant OMem thinks the hit is. A rough ranking, not a verdict — skim the summaries, not just the top score.
• kind — mail / file / calendar / loop, so you know what you’re looking at.
• page-id — a short handle (here a3f9c2bd) you pass to other commands.
• summary — the indented line: a one-line abstract, enough to judge relevance without opening anything.

To read the full page, pass its id (any 8+ character prefix works):

omem page get a3f9c2bd

It prints the wiki file verbatim — YAML frontmatter, then the curated body:

↻ replay

$

And if you need the original file itself (for a file hit), omem raw get <id> prints its path. That’s the progressive-disclosure ladder — summary, then page, then source — each step only when you need it.

1. omem ingest — build the wiki

2. omem query "…" — find the right page

3. omem page get <id> — read it in full

What’s next

You can query from the terminal — now let your AI agent do it automatically.

• Let an AI agent use your memory — Claude Code, Cursor, and others.
• Want more query filters (by date, kind, account)? Query and search.