I kept seeing the same architecture debate framed as a binary choice: should an agent remember through files or a vector database? Building Folderbase made me reject the premise. Files and folders are an authoritative working surface. Vector indexes are a retrieval mechanism.

Confusing them leads to brittle systems: either an unstructured folder the agent must repeatedly scan, or an opaque semantic store that cannot faithfully reconstruct the original work.

Files are where work remains usable

A file preserves its native form. A spreadsheet can retain formulas, a repository can retain executable source, an image can retain pixels and metadata, and a database file can remain queryable by its own engine. People and tools can continue working with those artifacts directly.

Folders also provide a universal namespace. Agents are already effective at listing directories, reading selected ranges, invoking local tools, and writing outputs. This makes the filesystem a strong execution surface and a durable source of truth.

Vector databases are indexes, not authorities

Vector search is valuable when an agent needs to find conceptually related material without knowing its exact path or wording. Embeddings can narrow a large corpus to relevant candidates and improve context selection.

But a vector record is usually a derived representation: a chunk, embedding, summary, or metadata row. It may be stale, incomplete, or unable to reconstruct the source. It should point back to an identified version of an authoritative object rather than silently become that object's replacement.

A folder database connects truth and recall

The missing layer is a durable model between the folder and its indexes. Stable object identity lets a semantic result resolve to the right file even after a rename. Version identifiers make retrieval reproducible. Provenance explains where a summary came from. Permissions filter candidates before sensitive content reaches an agent.

Folderbase treats semantic retrieval as one query capability among many. Exact metadata, relationships, lifecycle state, file type, ownership, and version history can work alongside embeddings rather than being flattened into them.

  • Files: authoritative bytes and native-tool compatibility
  • Folderbase protocol: identity, policy, versions, and relationships
  • Indexes: fast lexical, structural, or semantic discovery
  • Agent context: a bounded view assembled for the current task

Use the smallest context that can do the job

A good agent workflow does not upload an entire folder into every prompt. It queries structure, retrieves likely objects, checks permissions and versions, then materializes only the files or ranges needed for the task.

That approach reduces token waste without hiding the source. The agent can cite an exact object and version, while a person can open the same file and inspect the result.