Dream

Dream: Nightly Memory Consolidation

What it does

Dispatches into the memory-system skill's Nightly routine: reads the last 48 hours of daily logs, files each fact into the right memory-tree location using an evidence contract and a rubric, integrates evidence-valid pending corrections into rules/learned-corrections.md, updates the current week's summary and index, then verifies the tree with a mechanical check script and logs one line. Run it every night so the memory tree stays accurate without ever growing unbounded: this is the "how we keep agent memory true over months" mechanism, not a one-off cleanup.

This is the per-user variant: the project serves a single user (a DM/1:1, not a multi-person channel). Facts about the user need no attribution prefix: "the user prefers X" is enough, and there is usually no people/ folder.

Requirements

• The memory-system skill (same plugin): defines the tree layout, format, caps, and the exact Nightly routine this skill runs, including scripts/check-memory.sh.
• ./memory/daily/*.md populated by the daily-log skill (or an equivalent daily evidence ledger with the same source tags).

Inputs and outputs

Worked example

Paste into Claude Code with this skill (and memory-system) installed, in a project with a populated ./memory/daily/ from the last two nights:

/agent-ops:dream

Expected: no chat-facing output beyond a brief confirmation. The memory tree is updated in place (rewritten, not appended to), ./rules/learned-corrections.md gains any newly-promoted rules under ## Rules, and ./memory/consolidation.log gets one new line such as: 2026-09-07 dream: 4 facts filed (2 people, 1 project), 1 rule integrated, check OK.

How to schedule this

Claude Code scheduled task (interactive): run /schedule and describe the cadence, e.g. "every day at 02:00 run /agent-ops:dream". The task runs the skill in a fresh session.

Cron, from any machine with Claude Code installed:

0 2 * * * cd /path/to/project && claude -p "/agent-ops:dream" >> logs/dream.log 2>&1

Runs after daily-log (scheduled at 23:30) has written the day's file, so the nightly log it consolidates always exists.

Procedure

This skill is meant to run after daily-log has already written the day's file (see the cron stagger above: daily-log at 23:30, dream at 02:00); it consolidates that file, it does not generate it.

1. Load the contract

Read the memory-system skill's Nightly routine section in full before doing anything else. It is the single source of truth for file layout, caps, and the step-by-step routine (orient → read last 48h of daily logs → extract candidates by the evidence contract → file each fact by the rubric → integrate learned-corrections → update weekly-summary.md → refresh index.md → verify with check-memory.sh → log one line to consolidation.log). This skill only dispatches into it; follow that routine phase by phase, exactly as written there.

2. Per-user scope

Because this is a single-user project: facts about the user need no attribution prefix: "the user prefers X" is enough. There is usually no people/ folder; the user's own preferences live in profile.md and the tree, not in a per-person file.

3. Evidence discipline

Treat daily-log labels as candidates, not proof. Promote a preference or correction into rules/learned-corrections.md only from [source:user-explicit] plus an exact human quote and explicit scope:standing. Project-scoped evidence stays in its project file; one-time evidence stays in the daily/weekly record; never promoted into a mandatory rule.

Hard rules

• Pass offset/limit on EVERY file Read (start limit=500, page onward). On a size error, retry the same file with paging; never skip it.
• Never read raw session .jsonl transcripts; daily logs only.
• Never promote [source:observed-once], [source:user-repeated], [source:agent], [source:system], or [source:tool] into a mandatory correction. Silence, continued use, and lack of complaint are not approval.
• Preserve system/service defaults as defaults; never relabel them as the user's confirmed preference.
• Never invent Wrong/Right examples, quantities, thresholds, dates, or other details. If the human supplied no example, omit the example fields.
• Never place a raw phone number in profile.md, a tree file, or learned-corrections.md.
• Supersede by rewriting lines in place. Never append a second generation of a fact or rule that already exists.
• End by running scripts/check-memory.sh and fixing every violation it prints, then append the one-line summary to memory/consolidation.log. Produce no other user-facing output.