Organizing memory

These guides describe the current version of Space, which you can use today at app.getspace.sh.

Your categories. Your projects. A place for each useful detail.

You define your categories and projects, and Space can use a decision model to assign new memories to them. You can inspect and correct every assignment.

Start with categories you recognize

Open Memory → Categories. Create a category such as Dev, Accounting, or Personal and add a short description. Describe what belongs there, with examples or exclusions when names could be ambiguous.

Dev: Software development, code reviews, automated tests, and deployments. Exclude invoices for software subscriptions; those belong in Accounting.

Categories describe a subject. A memory can belong to more than one.

Connect your projects

A project identifies the work a memory concerns. For example, a deployment rule might belong to the Dev category and the Atlas project.

Create projects from Projects, then select them in a memory’s organization settings. Use Favorites to keep project and document shortcuts in the sidebar.

Understand automatic assignments

When classification is enabled, new memories and text edits are assessed against the current categories and projects. The decision model evaluates each possible match; Space applies the results. It can leave a memory unassigned when the context is insufficient.

• Global means a lasting preference or rule that applies across projects. It is a scope, not a default category.
• Specific projects means the memory identifies one or more existing projects. Sharing a topic alone is insufficient.
• Unassigned means no scope was established. The memory can still have useful categories.
• Uncategorized means no active category is assigned.

Labels organize memory; they do not grant access or restrict what an authorized agent can recall. The claim’s text and source references stay intact. Choosing a category does not prove that a claim is true.

Make a correction that stays

Open a memory’s details and choose Edit organization. Select its scope and labels, then Save assignment. This pins the complete assignment, including an intentionally empty one. Automatic edits and rebuilds preserve the pin for the same memory claim.

Replace with automatic classification explicitly releases the pin and asks the decision model to classify again. If that request fails or exceeds the supported size, Space keeps the previous assignment and reports the problem.

Organize existing memories

Enabling classification or creating a category does not automatically relabel older memories. Open an unpinned memory and choose Classify now. Renaming a category updates its displayed name, but changing descriptions does not rewrite existing assignments.

Archiving a category removes it from future automatic choices while preserving history and existing assignments. Restore it from the same dialog when needed.

Use the category filter and More memory filters to narrow by scope or project. These filters apply to loaded memories or current search results. Use Load older memory to bring in more items; text search uses the existing server search.

Processing and current limits

For each classification, OpenRouter and the model provider it routes to receive the memory’s title and text plus the current catalog’s names, descriptions, and kinds. Space maps the results to saved IDs. This is additional processing of your content; read Privacy & data before using sensitive information.

When your catalog has 32 active entries or fewer, each memory is assessed against all of them. Larger catalogs, up to 1,000 entries, are first narrowed to the 24 entries closest to that memory. Each request also carries a bounded amount of text; inputs beyond these limits remain unassigned instead of using an incomplete catalog. Manual assignment remains available.

Classification runs when memory is written or explicitly reclassified. Search has a separate decision-model relevance assessment, with a text-model fallback where needed. A failed classification can leave a new memory unassigned; it does not make that memory disappear. When the decision model was unavailable, a later change of model sorts those memories again: your next memory write starts one background pass that classifies only the memories left unassigned that way or by the limits above.