dream

Mem0 OSS Dream - Manual Memory Maintenance

Manual trigger only: run this skill only when the user explicitly invokes /mem0:dream or directly asks to run Dream memory maintenance. Do not infer a Dream run from memory count, noisy search results, elapsed time, routine health checks, or unrelated memory work. Do not schedule or start it in the background.

Fetch the active project, analyze everything in memory, show the proposed diff, then apply only after confirmation. Do not perform destructive changes before the report is shown.

Run steps in order. Do not skip the fetch completeness checks.

Modes

• Default / manual mode: read all memories, print the diff report, wait for user confirmation, then apply approved changes.
• --dry-run: read and report only. Never call add_memory, update_memory, or delete_memory.

Step 1: Resolve Scope And Policies

Use the active user_id and app_id from the environment or current Mem0 instructions. For this plugin, app_id is the project scope and must be passed as top-level app_id on writes, not only in metadata.

Load retention policies:

python3 "<PLUGIN_ROOT>/scripts/parse_mem0_config.py" "<cwd>"

If the parser fails or returns {}, use these OSS defaults:

metadata.type	Default retention
session_state	90 days
compact_summary	90 days
debugging_note with confidence < 0.5	180 days
all others	no pruning

Pinned memories (metadata.pinned == true) are never pruned or merged.

Step 2: Fetch All Project Memories

Call get_memories using explicit scope filters:

get_memories(
    filters={"AND": [{"user_id": "<active_user_id>"}, {"app_id": "<active_project_id>"}]},
    page=1,
    page_size=500,
)

Paginate until all pages are read. Continue while page * page_size < count. If the response contains truncated: true or complete: false, stop before analysis and report:

Dream paused. Memory listing is incomplete at fetch_limit=<N>; raise MEM0_OSS_LIST_FETCH_LIMIT or backend list limit before consolidation.