Memory

Anvil remembers. Not just within a conversation — across sessions, so the next time you work it already knows who you are, how you like things done, and the shape of your project. You stay in control of what it keeps.

The seven-layer model

Memory isn't one bucket. Anvil organises what it knows into layers, from the fleeting to the durable:

• Working / episodic — the current task and recent sessions.
• Semantic — durable facts: who you are, your confirmed preferences, project context, references.
• Procedural — repeatable steps and workflows it has learned.
• Identity & policy — the rules it operates under.

Each memory is a small, named note with a description, so it's inspectable — not an opaque embedding.

Curating it on AnvilHub

The Memory surface in your command centre shows the whole graph: every entry grouped by kind (About you, Your guidance, Project context, References, and more), each with its scope (this project vs. all projects) and how long ago it was learned. Expand any entry to read it — and forget one you no longer want kept.

[!NOTE] Your memory lives on your Anvil, never on the hub. The Memory surface in your command centre reads the graph over the relay — you can read what Anvil has learned and forget anything you want gone.

From the terminal

The /memory command gives you the same control in the session:

• /memory show <tier> — list what's held, by layer.
• /memory search <key> — find a specific entry.
• /memory promote <id> — make a nominated insight durable.
• /memory forget <key> — remove an entry.

It stays yours

Memory is held on your Anvil. The hub never stores or sees a memory entry — the Memory surface asks your Anvil for the graph over the relay and renders it; forgetting an entry deletes it on your Anvil. Nothing leaves the machine.

Related

• The workspace — the memory rail you see while working.
• Knowledge base — the durable, queryable library of your documents.