Memory and security for personal AI agents like Muse and Instinct
Personal AI agents such as Meta's Muse and Instinct keep a long-term memory of you, built from your chats, the patterns they notice and the accounts you connect, and they use it to act for you. That memory lives inside each company's service, and it is only as safe as the controls that company gives you.
Before trusting any agent's memory, ask: can you see and edit it, who can change it, is there a record of every change, where is it sent, and can you really delete it. Morphika is memory built to answer those questions, for the agents that can connect to it today: Claude Code, Codex and Claude Desktop.
About Morphika: Morphika is secure, shared memory for personal AI agents like Claude Code, Codex and Claude Desktop. How it is secured.
What is a personal AI agent?
A personal AI agent is an assistant that acts for one person: it reads your email and calendar, books, buys and replies, and keeps working between conversations. 2026 brought the first mass-market ones. Meta launched Muse on September 8, 2026, in the US, on iOS, Android and muse.ai. Instinct, from Spear Street Technology in San Francisco, is an invite-only agent you reach by text message or phone call.
Developer agents such as Claude Code and Codex are personal agents too: each works for one person, on that person's code and accounts, and each keeps its own memory.
How do Muse and Instinct remember things?
| Meta Muse | Instinct | |
|---|---|---|
| What it remembers | What you tell it, preferences and patterns it observes, and context from connected services, in a file-based memory system (Meta Help Center) | Everything happens in one ongoing conversation, and it runs its own persistent machine with browser access and cached credentials (Vellum) |
| See and edit | Ask Muse, or open Assistant → Identity → Memory; you can edit and ask it to revert changes | Not documented publicly |
| Checks before acting | Before sensitive actions such as sending an email or making a purchase (Meta) | Launch-week reports include an email sent without checking first (TechCrunch, Vellum) |
| Training | Using interactions to improve AI is on when you first use Muse and can be changed (Meta Help Center) | Terms grant a broad license over user materials, including for model training; the opt-out applies going forward (Vellum) |
| Deletion | Delete chats and files, ask it to forget, or Reset Muse. Meta notes Muse may still remember what it learned from deleted items | Reports of data kept after an account was disconnected (TechCrunch) |
| Outside memory (MCP) | Reviewed directory connectors, plus custom connectors (server address and OAuth client ID) that Meta says it doesn't review | No public way to add MCP servers found |
Both products are new and changing weekly. Everything in this table comes from the linked sources, checked on the date below.
What should I ask before trusting an agent with my memory?
- Can I see every memory and who or what wrote it? A memory you can't inspect is one you can't correct.
- Who can change it? An agent that reads your email can be steered by an email. Memory it writes on its own is an injection target.
- Is there a record? A tamper-evident log shows what changed, when, and through which connection.
- Is it used for training, and can I opt out?
- Does deleting really delete? Including what was learned from deleted items, and data from disconnected accounts.
- Can I limit and revoke access per connection, without shutting everything off?
How does Morphika answer those questions?
| Question | Morphika |
|---|---|
| See every memory | Yes, in the dashboard and through the agent's tools; each entry records which agent wrote it |
| Who can change it | Your personal memory: you and your connected agents. Shared memory: only after another member approves |
| A record | Every call, in a SHA-256 hash-chained audit log you can verify |
| Outside AI services | Search by meaning runs on an embedding model on Morphika's own server, so memory content and queries are not sent to an outside AI service |
| Deletion | Delete your account and personal memory yourself, from the iPhone app or the dashboard |
| Limit and revoke | A read-only address for untrusted agents, and per-connection revoke |
Can Muse or Instinct use Morphika today?
Not confirmed yet. Morphika connects through MCP with a browser sign-in (OAuth), and has no API keys by design. As of September 25, 2026:
- Muse: its custom connector form takes a server address and an OAuth client ID, which fits how Morphika signs in. Connecting the two is being tested now, and Morphika is not in Meta's reviewed directory. See memory for Meta Muse.
- Instinct: we found no public way to add an MCP server. See Instinct and memory.
The agents that connect today are Claude Code, Codex and Claude Desktop: see one memory for all your agents.
Sources
- Meta: Introducing Muse (September 8, 2026)
- Meta Help Center: Customize Muse's personality and memories
- Meta Help Center: Manage your Muse data
- TechCrunch: Instinct's AI assistant raises privacy and security concerns (August 24, 2026)
- Vellum: Official Instinct breakdown
- CellCog: Muse connector platform (September 19, 2026)
- Parallel: Meta Muse custom integrations
Sources checked September 25, 2026.
Related guides
Published September 25, 2026 · Last updated