Memory

Mobile Agent keeps a persistent memory so useful context survives between conversations. When you tell it something like "my project uses pnpm and targets API v2", it can record that as a fact and recall it later — without re-explaining everything each time.

The memory model

Memory is modeled as declarative facts, not chat transcripts:

• A fact is a short, self-contained statement (for example The mobile-agent build uses EAS.)
• Facts are stored in a memory.md document.
• The agent can add, update, and delete facts with the memory tools.
• Every change is recorded as a memory event (created / updated / deleted) with a reason and previous content, so edits are auditable.

This keeps memory compact and reliable — the agent stores what is true, not what happened.

How the agent uses memory

The memory tools are exposed to the model alongside the file tools. The agent is guided to:

• Store durable facts (preferences, project context, credentials-locations, conventions) when they surface in a conversation.
• Retrieve / search memory before answering questions that might depend on past context.
• Update facts when things change, rather than piling up duplicate entries.

You can see exactly what memory is stored and edit it manually from Settings → Memory.

Storage modes

Memory uses the app's storage abstraction:

• Local — the default. memory.md lives in the app's on-device database (SQLite via Drizzle). Nothing leaves your phone.
• Remote — optionally, memory can sync to a remote database endpoint configured in Settings, so the same facts are available across multiple devices.

Switching modes is done in Settings; existing local entries carry over into the active store.

Privacy

Memory is sensitive by nature. A few guarantees:

• Local memory never leaves the device.
• Memory content is only sent to the model currently active in a conversation that has memory enabled — never broadcast across providers.
• If you delete a fact or disable memory, the entry (and its history) is removed from the active store.

Example

1. Chat: "Remember that I prefer TypeScript and pnpm."
2. The agent calls the memory tool and stores the fact.
3. Week later, new conversation: "Set up my project."
4. The agent consults memory, sees the TypeScript/pnpm preference, and scaffolds accordingly — without you repeating it.

Note: Memory complements — and is distinct from — the on-disk file workspace. Workspace files hold long-form content the agent works on; memory holds compact facts the agent should always remember.