Task-focused recipes for learning how to use the AgentM harness. For field-level detail like flags and config keys, see Reference.

How-to	What it does
Install machine-wide (recommended)	The recommended setup — user scope, hooks, Drive vault.
Your first install	Fresh clone to a working scratch project.
Your first Agent View dispatch	The one-time auth setup a new machine needs before background dispatch works.
Configure a new project	First-session detect, propose, approve, then persist.
Re-detect a configured project	Re-scan a configured repo and diff against the stored config.
Update an installed harness	Pull a newer AgentM version onto your machine.

How-to	What it does
Run without a vault	Run the harness with no vault configured.
Choose a storage backend	Pick the storage backend, confirm its plugin.
Back the vault with Google Drive	Sync the vault across devices with Drive.
Set up Obsidian on the vault	Add the optional Obsidian layer over the vault.
Find your way around the vault	The seven root spaces and what each holds, the memory root, and the four checks that confirm every tool agrees.
Run the memory-root trims migration	Move an existing vault onto the post-plan-05 shape — standards/, Projects/agentm/, and the engine's own files out of the vault.
Run the card backfill	Bring every surviving note in the class directories to the card's shape — locked field order, required fields stamped, retired fields dropped, counter slugs renamed.
Run the maps and root notes migration	Retire Home.md and Filing.md for the generated root map, collapse numbered map pages into pages that paginate in place, and map each calendar year.
Run the root casing migration	Rename the four vault root spaces to their lowercase form — agent, calendar, personal, projects beside standards — to match the code that already spells them that way.
Run the projects migration	Dissolve every project's _harness/ into tasks/NNN-<verb-slug>/ directories, trackers written or moved, and the rest of the harness folder into the project skeleton.

How-to	What it does
Use AgentMemory in any agent	Read the vault from Claude.ai, Gemini, or Antigravity.

How-to	What it does
Use auto-context in phases	Recall budgets, save modes, thresholds per phase.
Tune auto-orchestration	Toggles and thresholds behind the briefing and idle chain.
Audit the vault	Run the read-only vault lint, apply fixes.
Read the morning note and the nightly scorecard	Read what last night did in the morning note, then the corpus scorecard in the order that makes it mean something.
Read the weekly field brief	Read the week's note on what is new in agent harnesses, memory, automation and skills, steer what it looks for, ask it a question, and keep the items worth keeping (pending).
Capture from your phone	Forward a link or a thought from the Claude app on your phone; it becomes a processed, recallable memory within about two sweep cycles.
Ingest an article	Turn a web page or a local file into a searchable memory — one intact full-document note plus reading-order-linked chunks.
Enable the daily email	Opt in to a daily email carrying the morning note, whose headline the SessionStart brief shows.
Enable the on-device notification	Opt in to a once-daily native notification with the brief's headline, for days you never open a session.
Persist a morning report to disk	Write an overnight run's morning report to a file, so it survives the terminal that produced it.
Review flagged memories	Work through the needs-review page's duplicates, contradictions, unfiled captures, low-confidence notes, and the dream cycle's twins and proposed facets.
Find missing note links	Discover and add the missing note links.
Tune the archive	Understand what ages a memory, where the thresholds and the cap live, and how to archive or revive one.
Archive a finished project	Retire a finished project's vault directory without losing anything load-bearing.
Use the daily calendar	Record a diary line or facet entry for today, and correct one logged wrong on a closed day.

Reference · Architecture · Home