Workspace memory is durable context — facts, decisions, and preferences that FilepadAI and connected agents should treat as ground truth without you repeating them every time. A well-curated memory makes agents more consistent and easier to debug. A vague or bloated one makes them worse than no memory at all.
What belongs in memory
A memory item is a specific, durable fact:
- Project-specific terminology or naming conventions
- Decisions that have been made and shouldn't be revisited
- Standing preferences (
output_format: markdown,language: formal) - Facts about the client, project, or domain that don't change run to run
What doesn't belong: anything that's only true for one run, anything you'd want re-derived from the source documents each time, or anything too vague to act on.
Shared vs. private
Memory items have a scope — shared (visible to the whole workspace) or private (visible only to you). Keep genuinely durable, agreed-upon facts shared; keep working notes and personal preferences private. You can promote a private item to shared once it's confirmed useful to the team.
Status: active, proposed, retired
An item can be active (in use), proposed (suggested by an agent, not yet confirmed), or retired (no longer current, kept for history rather than deleted outright). When FilepadAI or a connected agent learns something durable during a run, it can propose a memory item rather than writing it directly — you decide whether it becomes active.
Confidence and provenance
Each item can carry a confidence level and a note on where it came from — a specific document, a conversation, a manual entry. When something looks wrong later, provenance is what lets you trace it back to its source instead of guessing whether it was ever right.
Keep it small and explicit
A memory file with twenty precise, current facts beats one with two hundred stale ones. Review memory periodically:
- Retire anything no longer true
- Promote private items that turned out to matter to the whole team
- Watch for items that contradict each other — that's usually a sign a decision changed and the old memory was never retired
Checking what's in memory
Use the memory timeline to see what's been added, changed, or retired and when. If an agent's output seems to rely on outdated context, the timeline is usually the fastest way to find out why.
Next steps
- Core concepts — how memory fits into the workspace model
- Reviewing diffs and approvals — how proposed memory changes get reviewed