memory-status

Report the state of the Pseudolife memory stack:

memory_stats needs the core tool tier and memory_dream the full one: if your client does not list them, call memory_toolset(action="expand") (one tier per call) until it reports current: "full", then rediscover tools.

1. Fetch http://127.0.0.1:8765/health (curl or WebFetch). If it fails, report that the daemon is down and how to start it: docker compose -f <clone>/ops/docker-compose.yml up -d (install guide: https://github.com/Pseudogiant-xr/Pseudolife-MCP#quickstart).
2. Call memory_stats() and summarize: total_memories, store occupancy as a capacity meter (bands[0].size / capacity under the default flat preset — multi-band presets get a per-band table instead), the preset, and capacity pressure. A non-null capacity_warning means the store is within 20% of the point where every new memory permanently deletes an old one: lead the report with its message. true_drops_total counts those deletions all-time (last_true_drop names the most recent; true_drops counts only since the daemon started) — flag any non-zero. Then the reference bank (reference_bank_size / reference_document_count), and communities. Flag weights_reset: true if present — it means the store's counters restarted fresh. If read_audit is present, note its entries.never_read_pct as a coverage signal — the old-and-unread tail, not the young entries that are expected to be unread yet.
3. Call memory_dream(action="status") for consolidation state: backlog, idle_seconds, would_fire, and any pending outcome inference. backlog is what is still pending; dream_cursor is display metadata only and does not mark what has been consolidated. Fact/lesson counts and dream timing live here and in the Console, not in memory_stats().
4. If /health reports degraded (or any component error), surface the failing component verbatim — do not summarize it away. The same goes for a dream_tracking_error field: status stays ok, but the dream cannot pull or commit until it clears, and memory_dream(action="status") carries the detail.
5. Report the backup from /health's last_backup: its age_hours, and flag it when older than 36 hours (backups have stopped) or when rotation is held (entries, facts or lessons fell sharply, so the backup script kept every older copy — check for a wipe before anyone re-runs it with -AcceptRowDrop). If the key is absent, no backup has been recorded for this daemon: on the Docker tier, suggest ops/backup.ps1 and the daily ops/install-backup-task.ps1.
6. Mention the Cortex Console for browsing: http://127.0.0.1:8765/ui/