The AI assistant needs a file counter, not a memory palace

Durable assistant work depends less on mystical memory and more on plain files, clear folders, receipts, logs and handoff points that humans can inspect.

TL;DR

The reliable assistant does not remember everything. It writes things down in places the human can see, move, back up and delete.

What changed

The early assistant story was built around chat history: a long context window, a memory feature, a personalised model that “knows you”. Useful, yes. Sufficient, no.

Real work produces artefacts. Drafts. Spreadsheets. screenshots. PDFs. logs. invoices. half-finished plans. weird exports. things named final-final-v3. The assistant that only lives in chat eventually becomes a beautiful amnesiac with opinions.

A better assistant has a working filesystem and a shared exchange surface. Not because files are elegant, but because files are legible. Humans understand folders. Operating systems back them up. Other tools can open them. Permissions can be checked. A file can be attached to an email, dragged into a browser, compared in git, or thrown away.

Why it matters

For knowledge work, “memory” is often the wrong abstraction. The better abstraction is state.

Memory is what the assistant can recall. State is what the system can prove.

A durable workflow leaves receipts:

• source files used;
• drafts produced;
• decisions made;
• credentials not exposed;
• tests or checks run;
• current blockers;
• next actions;
• what changed since the last pass.

That is not glamorous. It is how you avoid the assistant saying “done” when the file was never saved, or summarising a plan that only existed in a context window that just compacted itself into soup.

The practical shape

The simplest useful pattern is a shared file counter:

Exchange/
  Inbox-from-human/
  Outbox-from-assistant/
  Shared-working/
  Archive/

The human drops things into inbox. The assistant writes outputs to outbox. Shared-working holds living project material. Archive keeps old runs from polluting the current task.

Add a short README.md at the top explaining what belongs there, and you have something most “agent memory” products still struggle to provide: a shared operational reality.

What to avoid

Do not share the whole home directory. Do not mix secrets with drafts. Do not let every automation dump into the same folder. Do not rely on filenames alone for sensitive material. Do not treat the file share as a replacement for permissions.

The goal is not to create a junk drawer. The goal is to create a handoff surface.

Watch next

The interesting agent platforms will make this boring:

• scoped file shares;
• human-visible run folders;
• automatic manifests;
• clean attachment handling;
• safe redaction flows;
• reversible deletes;
• search that distinguishes evidence from generated output.

The future assistant may be clever. Fine. But it also needs somewhere to put the damn files.