Rung 1 — Why Amnesia Is the Problem

The win: you stop blaming yourself when the agent "forgets" — and start seeing exactly which of the four memory tiers failed.

Run this experiment

1. Ask your agent something about your business it answered well last week. "What's our pricing structure?" or "What did we decide about the launch date?"
2. Watch it either guess, hedge, or confidently tell you something different than last time.
3. Don't fix it yet. Just notice the feeling — you hired a brilliant colleague who wakes up every morning with no idea who you are.

That's not a model problem. That's a memory problem, and it has a specific architecture.

The four memory tiers

Think of it like your own head:

• Journal — what happened recently. The running notes a session keeps about itself. Free, automatic, short-term.
• Snapshots — the save-point. When a session dies mid-thought, this is what lets the next one pick up the thread instead of starting cold.
• Named memories — the things you deliberately decided to keep. "Every Monday I do the pipeline review." Filed under a name, retrieved by that name. This is /remember and /recall.
• Semantic search — everything else, findable by meaning. You don't remember what you called it — /search finds it anyway.

Every "my agent forgot" moment is one of these four failing. Small stuff that just happened? Journal. Lost a session? Snapshot. A standing preference it ignored? You never /remember-ed it. Can't find last month's idea? That's a search job, not a naming job.

The takeaway

An agent without memory isn't a colleague — it's a very articulate goldfish. For the next four rungs we build each tier, smallest to biggest, until your agents carry your context so you don't have to.

Next rung: the memory you already have and aren't using — journals and snapshots.