Quick scan the conversation for any NEW information not already in memory: - User corrections or preferences - Environment changes - New workflows discovered - Problems solved with lessons learned
Save to memory only if genuinely new and durable (skip if already captured)
Save to skill if you discovered a reusable workflow or solved a complex problem (5+ tool calls)
Stop responding - the system handles the reset automatically
When the user asks to "get back into" or "pull up" a prior session: 1. Search previous session/context using distinctive project names, domains, paths, and issue phrasing. 2. Reconcile retrieved history with current injected memory, any compacted handoff summary, and the live source of truth. 3. If the user says the work was “mid-flight,” “the session we were in,” or otherwise identifies an unfinished active task, resume that task from the recovered state. Do not answer with a generic “What would you like me to do next?” when the handoff already contains a concrete active investigation or release path. Inspect the original screenshot/URL/repository/live service first when accessible, then continue the next unresolved gate. 4. Reply with concrete status evidence: working directory, live service/domain, process/port, tests/build/deploy/commit status, changed files, and screenshots/artifacts if available. 5. Avoid vague “we're up to date” claims unless backed by retrieved context or fresh verification. 6. Do not save completed task progress to memory; use memory only for durable user preferences or stable project facts.
If the user says something like "look at what you made" after a reset and the artifact/source is not visible in the current conversation, do not guess from the most recent session. Ask for or inspect the actual source first: screenshot path, URL, repo path, app name, or direct thread artifact. Session search is secondary context, not proof of the current target. Recency matches can be dangerously wrong when several projects share the same repo or app name (for example Astral Hermes rebuild vs a later rulership-lab session). When the user corrects the target session, immediately drop the stale session context and re-search with the user's exact anchors.