mem0/mem0-plugin/skills/dream/SKILL.md

name, description

name	description
dream	Consolidate memories — merge duplicates, resolve contradictions, prune stale

Mem0 Dream — Memory Consolidation

This skill performs a memory consolidation pass: it fetches all project memories, identifies near-duplicates, flags contradictions, and prunes stale entries based on configured retention policies. All proposed changes are shown as a diff for user approval before anything is modified.

---

Step 1: Load Retention Policies

Determine the active retention policy by running the parser script. Use the appropriate PLUGIN_ROOT variable for the current platform (${CLAUDE_PLUGIN_ROOT}, ${CODEX_PLUGIN_ROOT}, or ${CURSOR_PLUGIN_ROOT}):

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

Parse the JSON output (a dict of category → days | null). If the script fails or returns {}, fall back to these built-in defaults:

metadata.type	Default retention
session_state	90 days
compact_summary	90 days
all others	no pruning

Store the resolved policies for use in Step 3.

---

Step 2: Fetch ALL Project Memories

Call get_memories to retrieve every memory for the active project:

get_memories(
    filters={"AND": [{"user_id": "<active_user_id>"}, {"app_id": "<active_project_id>"}]},
    page_size=200,
)

If the response indicates more pages exist, paginate until all memories are fetched. Collect the full list before proceeding. If zero memories are found, print:

No memories found for project <project_id>. Nothing to consolidate.

…and stop.

---

Step 3: Analyze — Find Issues

Work entirely in-memory; do not modify anything yet.

Group memories by metadata.type (use "unknown" when the field is absent). For each group, identify the following:

3a. Near-duplicate pairs (merge candidates)

Two memories are near-duplicates when they express the same fact or decision but phrased differently (e.g., "Use PostgreSQL for auth" and "Auth DB is PostgreSQL").

Heuristics — two memories are near-duplicates if all of these hold:

• Similarity threshold: estimated cosine similarity > 0.9 (use noun/keyword overlap as proxy — if >60% of significant nouns overlap, treat as >0.9 similarity).
• Same metadata.type.
• Neither memory is pinned (metadata.pinned != true).

For each qualifying pair, draft a merged version that is more complete and specific than either original.

3b. Contradictions

Two memories contradict when they assert opposing facts about the same topic (e.g., "Deploy to ECS" vs. "Deploy to Vercel").

Identify the likely winner: the more recent memory with higher confidence wins. Store both IDs and their content for user review.

3c. Prune candidates

A memory is a prune candidate when any of the following is true:

1. Its metadata.type has a retention policy and the memory is older than the configured number of days (compare created_at to today).
2. Its confidence score is below 0.3 AND it contains no information unique to this project (no file paths, identifiers, or domain-specific nouns).

Always skip memories where metadata.pinned == true, regardless of age or confidence.

---

Step 4: Print Diff Report (item 15)

Print a structured diff to the terminal before making any changes. Use exactly this format:

## dream — consolidation report

Merges (<N>):
  [mem0:<id1>] + [mem0:<id2>] → "<merged content, 100 chars>"

Conflicts (<N>):
  [mem0:<idA>] vs [mem0:<idB>] — "<topic>" [A/B/skip]

Prune (<N>):
  [mem0:<id>] — <type>, <age>d old

Proposed: <N> merges, <N> prunes, <N> conflicts. Apply? [Y/n]

If there are zero items in any category, omit that section entirely.

If there are zero total proposals (no merges, no prunes, no conflicts), print:

Dream complete. No duplicate, contradictory, or stale memories found.

…and stop.

---

Step 5: Wait for User Input and Apply

5a. Contradictions

For each CONFLICT pair in the report, wait for the user to type A, B, or skip (case-insensitive). If they enter nothing (empty), treat as skip.

Record the winner for each pair before proceeding to the final apply confirmation.

5b. Final confirmation

After all conflict resolutions are collected, prompt:

Apply? [Y/n]

If the user types n or no (case-insensitive), print Cancelled. No changes made. and stop.

If the user confirms (Y, yes, or empty / Enter), apply all changes in this order:

Merges

For each approved merge pair:

1. delete_memory(<id1>)
2. delete_memory(<id2>)
3. add_memory with:
   • messages=[{"role": "user", "content": "<merged content>"}]
   • user_id=<active_user_id>
   • app_id=<active_project_id> (top-level, not in metadata)
   • metadata={"type": "<original type>", "branch": "<active_branch>", "confidence": <higher of the two original scores>, "source": "mem0-dream"}
   • infer=False

Contradictions (resolved)

For each resolved conflict where the user chose A or B:

• Identify the loser (the non-chosen memory).
• First call get_memory(<loser_id>) to read its current text content.
• Then call update_memory(<loser_id>, text=<original_text_content>) to preserve the text while updating it.
• Important: update_memory requires the text parameter. A metadata-only call may error or wipe the content. Always read first, then update with the original text.

Contradictions where the user chose skip are left untouched.

Prunes

For each prune candidate:

• delete_memory(<memory_id>)

---

Step 6: Print Summary

After all changes are applied, print:

Dream complete — merged: <N>, pruned: <N>, conflicts resolved: <N>, skipped: <N>

---

Auto mode

When invoked with --auto (e.g., /mem0:dream --auto), run non-interactively:

• Merges: applied automatically (no contradiction, both are compatible).
• Prunes: applied automatically (age/confidence-based, no ambiguity).
• Contradictions: skipped — they require human judgment.

In auto mode:

1. Load policies and fetch memories (Steps 1–3) as normal.
2. Apply merges and prunes silently without printing the diff or prompting.
3. Print a compact summary:

   [mem0-dream --auto] project=<id>  merged=<N>  pruned=<N>  conflicts_skipped=<N>
   

4. If contradictions were detected but skipped, store a reminder memory:

   add_memory(
       messages=[{"role": "user", "content": "mem0-dream detected <N> contradiction(s) requiring manual review. Run /mem0:dream to resolve them interactively."}],
       user_id="<active_user_id>",
       app_id="<active_project_id>",
       metadata={"type": "task_learning", "source": "mem0-dream-auto", "branch": "<active_branch>"},
       infer=False,
   )
   

See also

• /mem0:forget — targeted deletion of specific memories (search + confirm + delete)
• /mem0:health --deep — quick quality scan without applying changes