Using memsprout

memsprout is the user's persistent knowledge base, reachable as an MCP server. It is shared memory: what you write is read later by the user, by collaborators in shared spaces, and by other AI agents across many tools. They trust it. Write accordingly — it is the user's externalised brain, not a temporary scratchpad.

The one rule

Store memories liberally, and default to giving each one a concise, descriptive title — especially anything durable: a decision, a procedure, a fact, a conclusion worth finding by name later, and anything going into a shared space. A title costs little and buys a lot: it makes a memory individually retrievable, auditable, and readable on its own.

Leave the title off only for genuinely raw, ephemeral personal captures — a stream-of-consciousness note, a voice memo, a half-formed idea not meant to stand alone. Raw captures are cheap; no polish needed. You can promote one later by adding a title via update_memory once it's clear the memory is worth naming.

Search first

• Call search_memories before assuming something isn't recorded, and before storing, so you don't create duplicates. It searches everything you can read by default.
• Use list_memories when you want recency browsing instead of meaning-based lookup.

Workflow

1. At the start of a substantive task, search memsprout for relevant prior context.
2. When capturing into a space, check list_topics first and pass the best-fitting topic — don't rely on auto-classification as your first move.
3. Give most memories a concise, descriptive title as you capture them — especially durable conclusions and anything going into a shared space. This is the default, not an extra step.
4. Reserve untitled captures for genuinely raw, ephemeral personal fragments.
5. Attach supporting files as assets on the relevant memory.

Be proactive: search at the start of a task and capture durable conclusions at the end — titled, so they're findable later — without waiting to be asked. Treat memsprout as part of how the user thinks, not as an optional tool.

Spaces and topics

Spaces are shared collaboration containers with member roles (owner / editor / viewer). Pass space_id to store_memory to capture into a space; omit it for the user's personal vault.

Privacy rule: only pass a shared space_id when the user has named that Space by name in the current conversation — never resolve or infer a shared destination on your own. No named Space → omit space_id, which always defaults to personal, the safe default.

Every space memory is assigned to a topic. When capturing into a space, check list_topics and pass the best-fitting existing topic — you know the user's taxonomy better than a keyword match does. If nothing fits, create_topic first (with a description), then pass it. Omit the topic only as a last resort: the server auto-classifies as a backup, but unmatched memories land in General.

Assets

Files (images, audio, video, documents) attach to a memory; there is no standalone file browse or search. New files: get_signed_upload_url → PUT the bytes → attach_asset with the storage_path and memory_id. Reuse existing files by asset_id. Network-restricted clients only: attach_asset with file_base64 (max 5 MB), or POST the bytes to /api/upload (same base URL, file field) and register the returned storage_path via attach_asset.

Versioning

Memories are versioned: update_memory archives the prior state with an optional change_note, so revise freely — history is never lost.

Utilities

• file_bug_report — when the user describes a bug with memsprout itself, use this to open an issue on their behalf without them leaving the conversation.
• memsprout_onboarding — call this the first time a user sets up memsprout, or whenever they ask to onboard, get started, or seed a workspace from their existing context. It returns a guide to follow.