Memory

Memory lets an agent remember things between conversations: that you prefer short answers, that your work is tracked under #work, that a project you keep asking about lives in a particular folder.

It’s on by default. Each agent’s system prompt note, AGENT.md, ships with a # Memory section. Delete that section and the agent stops using memory. There is no separate toggle. See The system prompt is a note.

Where memories are stored

Memories are notes in Agents/Memories/. Open the folder and read them, edit them, or delete the ones you disagree with. There is no opaque store and nothing to export.

The folder is global: remembered facts belong to you, not to one agent, so every agent with a memory section shares it. What’s per-agent is the instructions for how eagerly to read and record, which are the # Memory section of that agent’s Agents/<Agent Name>/AGENT.md.

How the agent finds them

The agent doesn’t have to go looking. The # Memory section of its prompt ends with an index of the folder, rebuilt every time the prompt is assembled: one line per note, with the note’s path and a one-line summary. When a question could depend on remembered context, the agent reads the notes whose line fits, silently, before answering.

The summary comes from a description property in the note’s frontmatter. Notes the agent writes get one; add one to notes you write yourself, and keep the value quoted so a # or a : inside it isn’t read as YAML syntax:

---
description: "Where work is tracked: the #work tag and the [[Projects]] note"
---

A note without a description is listed by its first heading, or failing that its filename.

Always-loaded notes

A note with always: true in its frontmatter is included in full, not just listed. That is for the few facts every conversation needs: who you are, how you like answers. The plugin seeds one such note, Agents/Memories/User.md, the first time the folder is empty. It is a default, not a rule: rename it, rewrite it, or delete it and nothing breaks. The plugin only knows the two properties.

Both parts are bounded so memory can never crowd out the conversation:

Part	Budget
Always-loaded note bodies, combined	3000 characters
The list, all lines combined	4000 characters
One line of the list	160 characters

Past the budget an always-loaded note is cut short with a marker, and the list ends with a count of notes not shown, which the agent can still reach with list_directory. Notes themselves have no limit; a long note costs one line until the agent opens it.

The whole block is fenced and labelled as recorded facts rather than instructions, so a note that happens to read like a command is treated as a note.

Writes here auto-apply

This is the one place the agent writes without staging changes for your review. Requiring approval for every housekeeping edit to its own scratch space would make the feature unusable, so the trade is scoped tightly: auto-apply inside Agents/Memories/, review everywhere else.

The rest of your vault is unaffected. See Agents.

Requires note-writing

Recording a memory is a note write, so it needs manage_notes, which the manage-notes core skill attaches. An agent without that tool, because no enabled skill attaches it or a tool override vetoed it, can still read its memory folder (listing and reading notes come with explore-vault) but cannot add to it.

In that case the assembled prompt ends with a note that no write tools are enabled and that the agent must not claim to modify notes, so it should tell you it can’t record something rather than pretend it did.

Pointers, not copies

The default instructions are opinionated about what gets stored, and it’s worth understanding why.

The vault is your long-term memory and the source of truth. The memory folder is short-term memory the agent governs. So when something already lives in your notes, the agent is told to store a pointer (the tag, the wikilink, the folder, the search to run) rather than a copy.

At answer time it follows the pointer and re-reads the live note. A copy would be a snapshot that silently goes stale. A pointer stays correct as you edit.

Full content is reserved for facts with no home in the vault: a preference you stated in conversation and never wrote down.

The agent is also told to read memory silently: to check before saying it doesn’t know, rather than asking permission to look in its own folder.

Memory or skill?

Memory is not the only place the agent can keep what it learns. The instructions draw one line, keyed on who acts:

• How you work goes to memory: who you are, what you prefer, where things live in your vault. It is written as plain facts (“User prefers short answers”), never as instructions to itself, because an instruction re-read in a later conversation could override what you are asking for then.
• How the agent should do a kind of task goes into the skill it used for that task: the steps that worked, the pitfalls, and your corrections. Those load only when that kind of task comes up again.

Nothing that only mattered in one conversation is written anywhere. Before it finishes a task, the agent is told to ask itself whether it learned something worth keeping, route it to one of those two places, and otherwise save nothing.

Customizing

The # Memory section is part of a note. If the agent records too much, tell it to record less. If you want it to keep a specific kind of note, say so there. Two placeholders live in the section and are filled in each time the prompt is assembled: {{memoryFolder}} at the top becomes the folder path from the Agents folder setting, and {{memoryIndex}} at the end becomes the index described above. Leave both as they are.

The shipped default remains available as a diff (Diff with default in the Agent editor, shown once your note differs), so you can see what you’ve changed and reset if an edit doesn’t work out.

Excluded from search

Like the rest of Agents/, the memory folder is excluded from indexing, search, and the graph. Memories don’t pollute your search results, and the agent reaches them through the index in its prompt and by reading the folder directly.