hmem

Update flow for hmem. Runs `npm update -g`, syncs skills, applies migrations, verifies hooks, shows the changelog. Use when the user asks to update/upgrade hmem, hmem-mcp, or its-over-9k (any language), or when the startup version-check flags a new release. Runs the npm update itself — don't assume it's already done.

Switch active project mid-session via load_project, then fix any misrouted O-entry exchanges. Use when the user says 'aktiviere/lade Projekt X', 'switch to project Y', or whenever you'd otherwise call load_project mid-session.

View and change hmem memory settings, hooks, sync, and checkpoints. Use on /hmem-config, or when the user asks why context is huge, how to change auto-save, how often checkpoints fire, or how to set up sync.

Merge all session checkpoint summaries into one final O-Entry summary. Run at end of sessions with 2 or more checkpoints.

Load specific context from hmem based on what is needed RIGHT NOW. Use when load_project output is not enough for the current question.

Curate an .hmem file (your own or foreign) — mark obsolete/irrelevant, fix titles, consolidate duplicates, repair broken links. **Requires the `hmem-curate` MCP server** (skill prompts the user to enable it on entry). Use whenever the user says 'aufräumen', 'memory aufräumen', 'Speicher aufräumen', 'hmem aufräumen', 'clean up memory', 'tidy up hmem', 'curate memory', 'consolidate duplicates', 'merge duplicate entries', 'fix broken links', 'kümmer dich um die Memory', or invokes /hmem-curate. Also trigger when `memory_health()` flags BLOCKER/WARNING issues, when a P-entry's load_project output exceeds 4k tokens (session-start noise check defers to this skill), or before any batch cleanup of L, E, D, P entries. Skipping this skill and editing memory directly bypasses health checks, severity triage, and obsolete-chain integrity — never curate without it.

Dispatch an isolated sub-agent that returns only [RESULT]...[/RESULT]. Use for any search, lookup, 'does X exist?', 'find Y', calculation, or isolated writing — never do these inline when dispatch fits.

Migrate O-entries to the v5.3+ project-bound 5-level structure (each O00XX bound to its P00XX). Use on 'migrate O-entries', when hmem-update flags a structure migration, or when you notice O-entries still in the old flat L2→L4→L5 format.

Create a new E-entry (bug/error) in hmem with the strict 5-level scaffold. Use when the user wants to log a bug, file a report, or document an error ('Bug loggen', 'log this error', 'Fehler eintragen'), or before any write_memory with prefix='E'.

Create a new P-entry (project) in hmem via create_project — handles schema, section setup, and O-entry linking. Use when the user asks to add/register/track a new project ('neues Projekt', 'P-Entry erstellen'), or before any write_memory with prefix='P'.

Add a new rule and place it correctly — decide between a cross-project R-entry and a project-specific subnode under the active project's Rules section. Use whenever the user says 'neue Regel', 'Regel hinzufügen', 'new rule', 'add a rule', or invokes /hmem-new-rule. Critical safeguard: project-specific rules placed as R-entries pollute the session-start Rules listing across every project — get the scope right BEFORE writing.

Load long-term memory from hmem via read_memory or load_project. Use when starting fresh with no L1 summaries, after /compact, when the user asks 'what do you remember', 'continue where we left off', 'was war der letzte Stand', or about a project/topic visible in L1 summaries. Covers search, prefix filter, find_related, stale detection, and cross-project read_project().

Dispatch a Haiku sub-agent to search hmem for relevant memories. Sub-agent returns matching entries as ID + one-line summary. Main agent context stays clean.

Pre-publish checklist for hmem: skills synced, version bumped, tests green, nothing forgotten. Use before npm publish or when the user says 'release', 'publish', 'push a release', or 'neue Version'.

Search hmem when the user references something without an ID — past conversations ('letzte Woche', 'remember when'), unknown proper names, definite articles assuming shared context ('the bug we had', 'das Schema von gestern'), or asks whether a schema/rule/decision is documented. Search first, ask second. Skip for explicit ID lookups like 'read P0048'.

Search hmem memory for things referenced without an ID. Trigger when user refers to past conversations ("letzte Woche", "gestern", "remember when", "we talked about"), mentions proper names unknown this session, uses definite articles assuming shared context ("der Bug mit X", "the issue we had"), or seems to assume you know something. If you think "the user assumes I know this" — search first, ask second. Combines full-text search with time window from phrasing. Trigger on casual/vague references. Do NOT trigger for explicit ID lookups like "read P0048".

Mandatory entry point for every Cortex session — invoke at conversation start, after /clear, and after any load_project call. All stable context (H-entries, projects, device, sync) is pre-injected by the hook — no read_memory(mode='essentials') needed. Surfaces pending git work, Next Steps + open T-tasks, and runs the O-entry routing check.

First-time install of hmem: MCP server, skill files, and auto-memory hooks for Claude Code, Gemini CLI, or OpenCode. Use on 'set up memory', 'install hmem', 'initialize hmem', or 'first-time setup'.

Template for sub-agents dispatched by hmem-dispatch. Receive one task, return ONLY the result. No preamble, no explanation, no sign-off.

Set up hmem-sync for cross-device memory sync. Use on 'sync einrichten', 'setup sync', or when /hmem-config detects hmem-sync is missing. Covers first-device setup, additional-device restore, and MCP auto-sync.

Meta-skill that establishes how and when to use hmem skills. Loaded at session start. Defines mandatory habits for memory, dispatch, and lookup operations.

Prep for /clear: save high-value knowledge and update project Next Steps, then prompt the user to clear. Use on /wipe, 'clear/wipe context', 'context is full', or 'start fresh'. Not for deleting hmem entries, clearing terminal output, or cleaning up nodes.

Hmem write protocol — picks prefix (L/E/D/P/N/R/I/H…), tree location, and tags, and detects duplicates before persisting. Use before any write_memory or append_memory call (skipping it creates duplicates and misplaced entries), or when the user says 'remember this', 'save this', 'log this', or invokes /hmem-write.