Patch release focused on memory UX and docs (single enable path, automatic carryover of earlier notes, clearer /flush and /dream feedback), a more resilient /learn report stage, and making Cursor any-repo registration opt-in by default.
Changelog
[changed] Memory is documented and configured as one system again: enable with [memory] enabled = true or GROK_MEMORY=1; an explicit [memory] enabled = false turns it off even when managed remote settings would enable it. The separate memory-v2 opt-in path is no longer the user-facing model.
[new] Notes recorded with earlier Grok Build versions are carried over automatically the first time a workspace is opened after updating (each prior section becomes a topic, or is appended under “From earlier sessions” when the name already matches). Earlier files are left in place.
[changed]/flush shows “Flushing memory…” on the status line and posts a scrollback outcome (for example “Memory flushed through turn 12”). /dream shows “Consolidating memory…” and reports how many observations merged into how many topics, that there was nothing to consolidate, or that another session is already consolidating. Automatic background flush/Dream/session-end saves stay silent.
[fix]/learn no longer hangs forever when verifiers finish but the report step never writes report.md / actions.json. After a configurable stall window (default 20 minutes), wait exits with report_stalled, the parent can present the reducer synthesis as findings, and offer a single resume that re-runs only the report agent from the existing journal.
[fix] The /learn report agent is given a hard work budget (bounded file reads and tool calls) so it writes a shorter report instead of burning huge token counts re-reading the same inputs; failed or empty report agents surface the synthesis path clearly.
[changed]cursor_worker.any_repo is opt-in and defaults to off (previously defaulted on). When enabled, it still registers the any-repo door that clones whatever repository a Cloud Agent brings.
[changed] Memory guide copy describes global vs workspace scopes, topics/ plus observations folded by /dream, the bounded per-session index, and /memory diagnostics (s / y) without the old dual-stack rollout narrative.
Under the hood
Carryover path for earlier session notes into the current memory layout (per-workspace, non-destructive to the old files).
/learn wait loop tracks map/reduce/verify phases more accurately (including mapper artifacts at the run root), detects report-stage stall after all three verifier files exist, and returns the latest synthesis path for handover.
Bundled learn skill/workflow and state helper updated together so parent session, wait exits, and report-agent instructions stay aligned.
Bundled content manifest refreshed to public-2026-09-15-r1.
Upgrade notes
Prefer [memory] enabled = true (or GROK_MEMORY=1) going forward; you do not need a separate memory-v2 gate in user config. Local [memory] enabled = false still wins over managed enablement.
If you relied on Cursor Cloud Agents cloning arbitrary repos via the any-repo door, set cursor_worker.any_repo = true explicitly — it no longer defaults on.
Long /learn runs that stall only at the report step can be resumed for report-only retry while the original session is still open; a full relaunch is not required for that case.
