Memory — Persistent Notes for Your General Chat

Memory is a small persistent notebook the general Macha chat keeps for you across conversations. When you tell Macha "remember I prefer German-formal replies" or "note that our escalation channel is #support-legal," it saves that as a memory. On every later general chat, Macha reads that memory and adjusts how it works with you.

The feature is scoped to the general Macha chat — the one you open when there's no @agent in your message. Chats scoped to a specific agent already have their own instructions doing this job, so memory doesn't apply there. The chat header shows a pill on every conversation that tells you exactly which state you're in.

How memory works

Every memory is stored as three fields:

• Title — a short label, up to 200 characters. Example: "German reply tone."
• Description — a one-line summary, up to 400 characters. This is what the chat sees in its always-on memory index. Example: "German-formal Sie, never exclamation marks."
• Full text — the complete note, up to 10 000 characters. Only loaded into the chat when it decides the memory is relevant to what you just said.

Only the titles and one-line descriptions travel with every message — the full body stays out of the prompt until the model actively fetches it. This lets you write a memory that runs a few paragraphs without bloating every chat.

Saving a memory

The simplest way: tell Macha directly in a general chat. "Remember that I lead the CS team at MyPoster." "Remember I prefer bullet-point summaries." "Note that our WISMO SLA is 48 hours." Macha saves it and shows a black Memory saved toast when the write succeeds.

You can also add one manually from Settings → Memory. Click Manage to open the memory browser, then Add memory. Fill in title, description, and full text — the character counter next to each label turns amber as you approach the limit and red at the cap.

Memory is explicit-only for now — Macha will never save something without you asking. Autonomous "want me to remember that?" proposals are on the roadmap, but out of the initial scope.

Editing and forgetting a memory

Two ways to change a memory:

• In chat — "Actually, change my German reply preference to also avoid semicolons" or "Forget the WISMO SLA note." Macha figures out which entry you mean from its title, calls the update or forget tool, and shows a toast (Memory updated or Memory forgotten).
• In Settings → Memory → Manage — click a row to expand it into an editable form with title, description, and full text. Save changes when done, or use the trash icon in the collapsed row to delete. Deletion opens a confirmation dialog so it's never a single misclick.

The memory pill in chat

Every conversation shows a pill in the top-right header that tells you what state memory is in for that chat:

• Memory on (neutral grey) — general chat, memory is enabled, under capacity. Macha reads your memory index and can save, update, or forget entries.
• Memory 12/15 (amber) — you're approaching the 15-entry cap. Macha will start suggesting consolidations or deletions.
• Memory full 15/15 (amber) — new saves will fail until you delete something. Macha will tell you and ask which entry to forget.
• Memory paused (grey) — you've turned memory off in Settings. No reads, no writes.
• No memory here (grey): this chat is scoped to a specific @agent. The agent's own instructions serve the same role, so your memory doesn't apply.

Hover any pill and a tooltip explains the state in one sentence. The pill refreshes automatically the moment a memory is saved, updated, or deleted.

Turning memory off

Go to Settings → Memory and flip the Enable memory toggle off. When memory is disabled:

• Nothing is deleted — your saved memories are preserved and come back the moment you re-enable.
• Macha won't see any of your memories in the general chat and won't offer to save new ones.
• The chat header pill switches to Memory paused so you always know the state.

Scope and privacy

Memory is per (user, organization). If you belong to three Macha workspaces, you have three completely separate memory stores — one for each workspace, never shared with teammates. Your memories are yours; no other user (not even admins in the same org) can read them.

Widgets (the Chrome extension and the Zendesk widget) share your memory when they're in general chat, since the widget still signs you in as your Macha user via the same email you use on the dashboard. Talk to Macha in a widget's general chat and it will read from and write to the same memory store you see under Settings on the dashboard. Widget chats that target a specific agent (via the widget's configured @agent) still fall under "No memory here", the same as agent-scoped chats on the dashboard.

Limits

• Maximum 15 memories per user, per organization.
• Per-memory field limits: title 200, description 400, full text 10 000 characters.
• Both the manage page and the chat surface an approaching cap hint at 12 memories so you have time to prune before hitting the wall.

What memory is not

• Not a knowledge base. Memory is short, personal, tone-and-context — not a place to paste an SOP or article. For that, use Data Sources.
• Not shared with your agents. A scoped agent has its own instructions and doesn't inherit your memory. Move a permanent workflow rule into the agent's instructions instead.
• Not billable. Memory reads and writes don't consume credits.