Three kinds of memory, kept deliberately apart

"Give the AI a memory" usually means one thing: a single store it writes to and reads from. That works until it doesn't — the store fills with a mix of permanent facts and last-week's noise, nothing decays, and you can't tell what to trust. The fix isn't a bigger store. It's recognising that "memory" is actually three different jobs, and keeping them apart.

Three stores, three jobs

• Operational memory — small, durable facts about how I work: who I am, my preferences, a project constraint, a reference. One fact per file, with an index, recalled at the start of a session before anything substantive happens. It's the layer that stops every session beginning from zero.
• The workshop — where the work actually happens: daily notes, decisions, in-flight project material. It's time-bound by design — most of it is useful this week and irrelevant next month, and that's fine. Crucially, it's indexed, so I can ask it questions in plain English and get answers cited back to the source file, instead of digging by hand.
• The library — a separate Obsidian vault of durable concepts: the things that stay true across projects, written as linked notes in a graph. It's small on purpose, and it lives in its own repository with its own rules — a tool built for connected notes, not chronological ones.

How they connect

Two flows tie the stores together, and both are deliberate:

• Retrieval. The workshop is indexed, so it answers questions — "what did I decide about X, and when?" — with the source attached. The memory becomes askable, not just stored.
• Promotion. At a weekly review I ask what, from the week's workshop churn, has earned a place in the library. Capture is wide; promotion is narrow. Capturing something creates a memory, not a commitment — and that distinction is what stops the durable layer filling up with noise.

Sitting across all of it is the operational memory, read at the top of each session so context compounds instead of resetting.

Why three, not one

This is the judgement, and it's the same instinct each time: the three layers answer different questions and decay at different rates, so they need different rules. The workshop is allowed to be messy and ephemeral; the library has to stay trustworthy; the operational memory has to be correct right now. Put them in one bucket and you get the worst of each — durable knowledge buried in transient noise, or transient notes over-curated as if they were permanent.

There's a second, sharper reason to keep the operational layer disciplined: a recalled memory is background, not a command. It reflects what was true when it was written, so if it names a file or a setting, that's re-checked against reality before anything is done with it. A memory treated as gospel quietly rots as the world moves on — and treating stored text as instructions is exactly how a memory layer becomes an attack surface. So the rule is the opposite: memory informs, reality decides, and the store is curated — wrong notes deleted, duplicates merged — not hoarded.

The two deep-dives

This is the map. Two of the three layers have their own write-ups:

• Giving AI a memory that lasts — the operational memory, and the discipline that keeps it trustworthy.
• Keeping what I learn — the workshop-and-library split, and why the boundary between them is the point.

What's shown here is the architecture: how the three stores fit, and the rules that keep them honest. The notes themselves, and the personal material in them, stay behind the boundary where they belong.