AI Memory
What it does: AI Memory lets you bring context from other AI assistants (ChatGPT, Claude, Gemini) into Cumulus. Your memories are organized into three pools — identity, instructions, and knowledge — so the AI personalizes every response without you repeating yourself.
How it works
Memories are split into three categories with different retrieval strategies:
| Pool | What it stores | How it’s used |
|---|---|---|
| Identity | Name, role, company, location | Always included in every conversation |
| Instructions | Tone preferences, format rules, “always/never” directives | Always included in every conversation |
| Knowledge | Projects, tools, skills, workflows, technical context | Semantically matched to your current question (top 8 most relevant) |
Importing memories
Copy the export prompt
After onboarding (or from Settings), the import dialog gives you a universal prompt. Click Copy Prompt to copy it.
Paste into your current AI
Open ChatGPT, Claude, Gemini, or any other AI assistant and paste the prompt. It will output everything it knows about you.
Paste the result back
Copy the AI’s response and paste it into the import dialog in Cumulus. Click Import.
Review the results
Cumulus automatically breaks large entries into individual facts, classifies each one, and skips duplicates. You’ll see a summary of how many were imported.
Pro tip: You can also add memories manually from Settings → AI Memory. Click Add and choose the category (identity, instruction, or knowledge).
Note: Each user can store up to 200 memories. Identity and instruction memories are always sent to the AI. Knowledge memories are only included when they’re relevant to your current question.
Managing memories
Go to Settings → AI Memory to:
- View all memories grouped by type (Identity, Instructions, Knowledge)
- Edit any memory inline
- Delete memories you no longer need
- Import more memories at any time
- Add individual memories manually
Admin controls
Organization admins can:
- Toggle AI Memory on or off for the entire organization from Settings
- Per-workspace toggle: Disable memory injection for specific workspaces in workspace settings
Tips & gotchas
- One prompt fits all: The same export prompt works for ChatGPT, Claude, Gemini, and any other AI — no platform-specific setup needed
- Deduplication: If you import the same memories twice, duplicates are automatically skipped (cosine similarity > 0.95)
- Atomization: Large multi-paragraph memories from ChatGPT are automatically broken into individual facts, each searchable on its own
- Privacy: Memories are per-user and protected by row-level security — no one else in your org can see them
Project Memories
Project Memories are separate from your personal AI Memory and scoped to a single project. They’re shared with everyone on that project’s team.
Use them to store context that should apply to every chat in the project — the goal of the work, who the customer is, glossary terms, output format preferences, or anything else the AI should keep in mind across all conversations there.
| Personal AI Memory | Project Memory | |
|---|---|---|
| Scope | You, across all chats | One project, shared by all members |
| Who edits | You | Project editors |
| What it stores | Your role, preferences, instructions | Project goals, customer context, glossary, format rules |
Who can edit: any team member with editor access on the project. See Member Invitations for role definitions.
How they combine: in a project chat, the AI has access to both your personal memory and the project memory at the same time. Personal memory tunes the AI to you — your role, your communication preferences. Project memory tunes it to the work itself.
Project memory is set from the Memory tab inside a workspace. Personal AI Memory is managed in Settings → AI Memory. They’re configured separately and applied together.