Memory

Memory is what makes an assistant feel like a teammate rather than a chatbot. Each assistant has its own memory; assistants don’t share. The memory survives across sessions, machine restarts, and engine updates.

What’s in memory

Memory has four parts, all stored as plain Markdown that you can read and edit:

Long-term

Typed facts the assistant chooses to remember. The types are:

• User — about you. “Prefers four-space indentation.” “Is a senior Rust engineer.”
• Project — about an in-flight initiative. “Auth rewrite freezes Friday.” “Database migration is gated by legal review.”
• Feedback — corrections or confirmations from you. “Don’t mock the database in tests.” “Yes, one bundled PR was the right call.”
• Teams — about a team the assistant belongs to (file prefix teams_).
• Organization — about the organization the assistant belongs to.

Each long-term entry is one file with a slug like user_indentation-preference.md. The assistant writes them itself, but you can edit them by hand.

Reference

Files you seed for the assistant. Things you want it to know but don’t want it to “discover” on its own — coding style guides, architecture overviews, deployment playbooks. The assistant treats these as read-only.

Journal

A daily log. One file per day. The assistant appends timestamped notes throughout the day; you can scroll back through what it was thinking last Tuesday.

Today and yesterday are loaded as part of every wake briefing. Older entries are searchable from the assistant but aren’t loaded automatically.

Index

An auto-generated table of contents. Rebuilt on every memory write so the assistant can find what it has stored. You don’t edit this directly. In the Memory tab the matching sub-tab is labeled History and also shows the per-file change history.

How the assistant uses memory

Every wake briefing includes:

• The index (so the assistant knows what it knows).
• Today’s and yesterday’s journal.
• Any long-term entries the assistant explicitly recalls.

The assistant adds to memory when something feels worth keeping: a correction from you, a project decision, a non-obvious fact. The assistant writes routine appends without asking; the more destructive operations (replacing or deleting an entry) require you to approve.

You can ask the assistant to remember something explicitly. “Remember that this codebase uses snake_case for filenames.” That ends up in long-term as a user or project entry.

The Memory tab

On desktop, open the assistant and click Memory. Four sub-tabs match the four parts above:

• Long-term — list of entries, grouped by type. Click one to edit.
• Reference — the files you’ve seeded.
• Journal — one entry per day, ordered most-recent first.
• Index — the auto-generated table of contents.

You can edit any file directly. Saves are versioned in the assistant’s local history, so you can undo changes from inside the app.

On the portal, the same four sub-tabs are present.

On mobile, Memory is a single screen with the same sub-tabs as buttons across the top.

Soft and hard caps

Total long-term storage is capped to keep wake briefings fast. You’ll see a warning when an assistant’s long-term hits 40 KB total; writes are blocked at 150 KB. Reference and journal don’t count toward those caps.

Most assistants stay well under the warning. The cap exists to keep wake latency predictable; if you’re hitting it, summarize old entries into newer, broader ones.

Versioning

Each memory write is recorded in a local version history just for that assistant. You can see who changed what and when from the Memory tab. Edits made by the assistant are tagged as assistant edits; edits made by you (or another user) are tagged with the user.

What’s separate from memory

• Persona and constitution — those are stable settings, not memory. See Persona & constitution.
• Channel transcript — chat between you and the assistant. Loaded as recent context, not as long-term memory. See Chat.
• Workbench — the assistant’s working folder for drafts and intermediate files. See Workbench.

Next step

• Workbench for the assistant’s working folder.
• Routines for scheduled assistant wakes.
• Assistant governance for permission tiers and budgets.