Unlike the guy in Memento, your AI actually remembers correctly — and knows when to forget.

Install · How It Works · Artifacts · Multi-AI · Obsidian

Memento OS is a persistent memory plugin for AI coding tools. Every conclusion is stored as a reasoning artifact with an invalidation trigger — your agent remembers what matters and knows when to re-evaluate.

AI agents are stateless. Every session starts from zero.

• Decisions discussed → gone when the session ends
• Context compressed → compaction kills memory
• Research done once → invisible next session
• "We chose Supabase" → nobody remembers why, or when that stops being true

/plugin install memento-os

Then:

/memento:init

Three questions. Thirty seconds. Your first decision captured.

Five skills. One loop.

Hooks run automatically — Stop captures artifacts on session end, PreCompact saves before compression.

Commands: /memento:init (setup) · /memento:stats (memory score + streaks)

The atomic unit. Not a note — a pre-computed conclusion with an expiration condition.

[D] Use Supabase over Firebase — invalidates if Firebase adds RLS            [critical]
[I] MV3 service workers die after 30s — invalidates if Chrome changes policy [settled]
[E] Deployed without testing webhooks — root cause: no staging env — fix: added pre-deploy checklist [settled]
[S] Consider caching layer — activates when: API p95 > 200ms                [volatile]

Confidence (Claude proposes) × Impact (you confirm):

Cap: 24 artifacts per context. Eviction: noise → settled → volatile → critical (never).

Same vault, same artifacts, different integration depth.

Memory is plain markdown. Point Obsidian at your vault for graph view, backlinks, and search.

knowledge-vault/
├── Projects/
│   ├── _context.md          # Cross-project decisions
│   └── my-app/
│       ├── _context.md      # Project artifacts (the brain)
│       ├── NEXT.md          # Session continuity
│       ├── Decisions/       # Full decision records
│       └── Sessions/        # Session log
├── Personal/
│   └── _context.md          # Hobbies, health, learning artifacts
├── Business/
│   └── _context.md          # Cross-business decisions
├── Knowledge/
│   ├── _context.md          # General cross-domain insights
│   └── MOC — Patterns.md    # Map of Content hubs
├── People/
│   └── Self.md              # Your peer card
└── _meta/
    └── conventions.md       # Vault rules

Every area has its own _context.md with an artifacts table. Artifacts route automatically to the right context based on domain. Tags: tech/, business/, product/, ai/, personal/ — nested as domain/subtopic. Folders created on demand.

Without (session 5):

> What auth approach did we decide on?
I don't have context on previous decisions. Could you remind me?

With (session 5):

## Session Briefing — My Project
Memory: 6.8/10 | 14 artifacts | 2 seeds | streak: 5

### Active Decisions
[D] OAuth via Supabase Auth — invalidates if rate limits hit [critical]
[D] Mobile-first, no desktop v1 — invalidates if desktop demand >30% [critical]

### Seeds Ready
[S] Consider Redis caching — activates when: API p95 > 200ms ← CONDITION MET

I kept reaching the same conclusions across sessions — same reasoning, same answer, different day. The vault was full of notes, but the actual decisions were buried in noise.

So I flipped it: capture only conclusions. Each one with a condition that makes it invalid — so the system knows when to question itself instead of blindly trusting old decisions.

That shift — from "save everything" to "save only conclusions" — turned out to be the entire product.

The /memento:grill-me skill is based on Matt Pocock's grill-me prompt, adapted for the Memento OS artifact system. If you like what it does, check out his repos — there's more where that came from.

Welcome: evolution entries, adapter improvements, scorecard improvements, translations.

Not accepted: vendor plugins, paid integrations, AI-generated filler.

If Memento OS saved you time:

paypal.me/aiyo28 — any amount, any currency.

MIT