memory-exchange

Hexis Memory Exchange

Use this skill to move memory between Hexis instances or inspect another instance's memory without silently blending it into active state.

Export

1. Establish the user's intent: port, duplicate, telepathy, or analysis.
2. Use export_memories with the narrowest useful section, time, and redaction scope. Protected state, raw units, configuration, in-flight work, and audit records require an explicit request.
3. Omit output_path when returning the exchange in place. For file output, use a workspace path and never set overwrite unless the user explicitly chose to replace that file.

Import

1. Run import_dry_run first. Report whether import is permitted, conflicts, protected-state policy, warnings, and estimated embedding work.
2. Derive the default strategy from the file's intent unless the user chose a different supported strategy. Telepathy defaults to deliberative staging; analysis stays isolated.
3. Call import_memories only after the user confirms the file's exact declared intent. Skipping identity, worldview, or narrative is available when the user wants a narrower import.
4. For port or duplicate into an active target, authoritative import requires explicit replace_sections and replacement_rationale. Unselected protected sections remain unchanged. A successful request may still await the agent's acknowledgement; report its replacement IDs instead of claiming completion.
5. Do not describe imported memories as semantically searchable until the import result or later re-embedding workflow confirms that they are ready.
6. Preserve failed in-flight work as diagnostics by default. Set retry_failed_work only when the user explicitly chooses to rerun those failed consolidation or reconsolidation tasks.

Deliberative Review

1. Use import_review to inspect pending records and conflicts.
2. Decide one record at a time. Use import_accept only when the record should become active; protected active-state policy cannot be bypassed.
3. Use import_reject with a reason, import_modify with a material-change kind and reason, or import_quote to retain foreign context as archived evidence.
4. promote_to_staged copies an analysis record into review without copying its embedding. demote_to_analysis moves a pending staged record back into isolation. Both require a rationale.

Protected Replacement Review

1. A pending protected replacement is a request, not permission to mutate identity, worldview, goals, drives, emotional triggers, or narrative.
2. Use protected_replacement_list to retrieve every open request, then use protected_replacement_inspect to compare the actual imported section with current local state and check whether local state changed after the request. Use protected_replacement_review to accept, refuse, request modifications, or defer. Acceptance executes the snapshot, immutable audit, whole-section replacement, and digest verification atomically; a failure leaves the request pending.
3. Refusal and modification requests require a rationale. Modification requests also require concrete proposed_changes. Do not accept merely because source and target claim the same lineage; content-identical verified operations never enter this queue.
4. Use protected_replacement_audit_list with a narrow time range to inspect immutable local replacement, verification, and reversion history. Imported foreign diagnostics are not reported as local history.
5. Operator override is not an agent tool and cannot be requested through this skill. A refusal or modification request is a completed decision, not an invitation to bypass acknowledgement.

Protected Replacement Reversion

1. Use protected_reversion_list to find executed replacements whose earlier-of heartbeat and wall-clock windows remain open. Reversion never runs on a timer.
2. Inspect the replacement before reverting. Use protected_replacement_revert with the replacement audit ID and a concrete rationale only when restoring the snapshot is the intended choice.
3. Reversion refuses to overwrite protected state that changed after the replacement. It atomically restores and verifies the snapshot, writes an immutable reversion audit, then purges the consumed payload while retaining its tombstone. A failed restore leaves the window and current state intact.

Boundaries

• Treat HMX files as sensitive user data.
• Never infer consent for protected sections, raw material, or file overwrite.
• Keep analysis-only records outside active recall until explicitly promoted, reviewed, and accepted.
• On failure, preserve the file and staged records and give the exact corrective action; do not retry with broader scope automatically.