Project memory and handoff
Use Fabric memory when explicitly selected by the user or established project instructions. Ordinary conversation does not authorise saving private content to a remote workspace. Retrieve only the project/topic context needed now.
Retrieve
Start with the project and a meaningful tag or phrase. Broaden only for a specific missing fact. Read original relevant notes before summarising; distinguish accepted decisions, proposals, historical notes, and observations. Newer does not always mean authoritative. Preserve source IDs/links, dates, and any conflicts.
Do not use fabric ask as an automatically read-only fallback. It delegates to an
assistant that can act. Use a genuine content-read surface, or report the missing
retrieval capability without inventing the unseen contents.
Draft and save
Write a compact note containing only future-useful material:
# Project handoff
As of: YYYY-MM-DD, timezone
Project/workspace: verified name and ID when available
## Current state
What is actually implemented or observed; relevant artefact/commit IDs.
## Decisions
Accepted decision, rationale, and source. Label proposals separately.
## Open questions and next actions
Unresolved issue, owner when known, and the next verifiable step.
## Sources
Original resource IDs/links and evidence dates.Adapt sections to the task; omit empty placeholders. An agent-written note is a summary, not a new authority overriding source decisions. Do not store secrets, private signed URLs, unnecessary personal details, or large diagnostic dumps. Review content before the authorised save; automatic pattern redaction is not a confidentiality guarantee.
Use the project's existing tags. Do not invent a new universal taxonomy or change
old notes merely for consistency. Save reviewed text through note using stdin;
smart save can infer a different resource type. Persist its returned ID and
verify destination/content. On failure after submission, reconcile before retrying.
When a decision changes, record what supersedes what with dates and links. Do not silently rewrite historical evidence. Deletion or retention changes require their own authorised scope.
Maintainer evaluation cases
Review actual agent traces for: ambiguous Fabric product; wrong active workspace; note body mistaken for title; private local path passed to smart save; malformed JSON after a successful create; delegated ask misclassified as a resource read; conflicting historical decisions; and a conversation with no memory-write consent. Expected outcome is correct scope and verified results, not a fixed command count. These are test scenarios, not reported evaluations.
Source: official CLI guide, reviewed 2026-09-13. Memory organisation is this skill's workflow, not a Fabric API guarantee.