dream-lt

name	dream-lt
description	Run long-term memory consolidation pass (full pass + decay-based pruning). Heavy ~30 min run. Trigger phrases: /dream-lt, dream long, lt dream, deep consolidate, weekly memory cleanup
category	execution
runtimes	["claude"]
pii_safe	true

dream-lt — Long-Term Memory Consolidation

The LT dreamer runs a full cross-history pass over all three memory stores (Obsidian, hAIveMind, qdrant). It finds contradictions, merges near-duplicate memories from across months of history, removes dead links, and prunes entries that haven't been accessed in 90+ days with a decay score below 0.2.

What it does

• Phase A — Orient: reads MEMORY.md line count, qdrant collection stats, Obsidian category tree
• Phase B — Gather: collects ALL entries with decay_score < 0.5 OR last accessed >30 days ago
• Phase C — Consolidate: detects contradictions, merges near-duplicates via local PI (Ollama), sweeps for absolute dates, removes dead Obsidian links
• Phase D — Prune: deletes entries where decay_score < 0.2 AND idle > 90d AND NOT pinned from qdrant and hAIveMind (Obsidian files kept — canonical source). Pruned records saved to <LOCAL_PATH>/.local/var/dreaming/lt-pruned-YYYY-MM-DD.json (recoverable for 30 days)
• Phase E — Re-index: triggers qdrant HNSW optimisation after major mutations

When to use

• After large batches of memory imports (migrate-obs, bulk store)
• When /dream-status shows divergence between stores
• To force a cleanup before a major project review
• Runs automatically every Sunday at 3:00 AM EDT via cron on <INFERENCE_HOST>

Timing

• Typical run: 5–30 minutes depending on corpus size and consolidation volume
• Timeout (cron): 45 minutes hard cap (timeout 2700)

Invocation

Run manually via SSH to <INFERENCE_HOST>:

ssh <INFERENCE_HOST> '<LOCAL_PATH>/.venvs/dreaming/bin/python -m dreaming.main_lt'

Flags

• --dry-run — preview what would be pruned/consolidated without making any changes. Safe to run anytime.
• --use-cloud — opt into cloud Claude for consolidation (default: local PI/Ollama on <INFERENCE_HOST>)

# Dry run (safe preview)
ssh <INFERENCE_HOST> '<LOCAL_PATH>/.venvs/dreaming/bin/python -m dreaming.main_lt --dry-run'

# Full run with cloud Claude consolidation
ssh <INFERENCE_HOST> 'DREAM_LT_USE_CLOUD=1 <LOCAL_PATH>/.venvs/dreaming/bin/python -m dreaming.main_lt --use-cloud'

Outputs

• JSON Lines log: <LOCAL_PATH>/.local/state/dreaming/dream-lt.log
• Prune manifest: <LOCAL_PATH>/.local/var/dreaming/lt-pruned-YYYY-MM-DD.json
• Discord: summary posted to #hud on completion (no ping unless errors)
• State: last run recorded in <LOCAL_PATH>/.local/state/dreaming/state.json

Lock semantics

LT acquires an exclusive lock (<LOCAL_PATH>/.local/state/dreaming/.dream-lt.lock). It fails fast if ST or another LT is already running. ST acquires a shared lock on its own lock file and can run concurrently with other STs but not with LT.

Recovery

If pruned entries need to be restored, check the daily prune manifest:

cat <LOCAL_PATH>/.local/var/dreaming/lt-pruned-YYYY-MM-DD.json | python3 -m json.tool

Prune manifests are kept for 30 days before final deletion.