Janbjorge Rekal Rekal

Find and fix problems in the rekal database. Every change requires explicit user approval. Never auto-delete. Never auto-modify.

Step 1: Health overview

memory_health()

Report one line:

"142 memories across 3 projects, spanning 6 months. 4 conflicts detected."

Step 2: Resolve conflicts

memory_conflicts()  # global first

If > 10 conflicts, also run per-project: memory_conflicts(project="<name>")

Per conflict pair, classify and propose:

Conflict type?
├── One outdated, one current
│   └── Propose: memory_supersede(old_id="<outdated>", new_content="<current content>")
│
├── Both valid, different scope (e.g. PostgreSQL for OLTP, ClickHouse for analytics)
│   └── Propose: Keep both. Add project scope if missing.
│
├── Genuine contradiction, unclear which is correct
│   └── Ask user: "Which is correct? [A] or [B]?"
│
└── False positive (not actually contradictory)
    └── Propose: Remove contradicts link.
        memory_link(from_id="<id_a>", to_id="<id_b>", relation="related_to")
        — or memory_delete the link if truly unrelated

Present as numbered list:

1. "API uses v2" vs "API migrated to v3" Proposal: Supersede v2 with v3. [approve/reject]
2. "Use PostgreSQL" vs "Use ClickHouse for analytics" Proposal: Not a real conflict — different use cases. Remove link. [approve/reject]

Step 3: Find duplicates

memory_topics()  # or memory_topics(project="<name>")

For every topic cluster with count >= 3, search for near-duplicates:

memory_search(query="<topic name>", limit=10)

Read results. Group memories covering the same fact/preference/procedure.

Per duplicate group:

1. Identify the most complete and accurate version
2. Propose superseding all others into it

Format:

Duplicates (formatting preferences):

• mem_abc: "User prefers Ruff" (2024-01)
• mem_def: "User prefers Ruff over Black for formatting" (2024-03)
• mem_ghi: "Use Ruff, not Black. Also handles import sorting." (2024-06)

Proposal: Keep mem_ghi (most complete). Supersede mem_abc and mem_def into it.

memory_supersede(old_id="mem_abc", new_content="<content from mem_ghi>")
memory_supersede(old_id="mem_def", new_content="<content from mem_ghi>")

Step 4: Stale conversations

conversation_stale(days=30)
conversation_threads(limit=20)

Flag conversations with 0 memories as cleanup candidates. Conversations with memories are fine — memories persist regardless.

Conversations are cheap storage. Only flag truly empty ones older than 30 days.

Step 5: Quality audit

memory_timeline(limit=20)

Skip memories created < 24 hours ago — too fresh to judge.

Per memory, check against these rules:

Quality issue?
├── Content < 20 characters
│   └── Propose: reword with more context, or delete if worthless
│       memory_update(memory_id="<id>", content="<expanded content>")
│
├── Project-specific content but project=None
│   └── Propose: add scope
│       memory_update(memory_id="<id>", project="<correct project>") — NOT SUPPORTED
│       Note: memory_update cannot set project. Propose memory_supersede with project= set.
│
├── Wrong memory_type (e.g. a procedure stored as fact)
│   └── Propose: fix type
│       memory_update(memory_id="<id>", memory_type="<correct type>")
│
├── context-type AND older than 60 days
│   └── Propose: still true? → memory_update to fact. Stale? → memory_delete.
│       Ask user which.
│
└── No issues → skip

Step 6: Present action plan

Compile ALL proposals from steps 2-5 into one summary:

Hygiene report:

• 4 conflicts: 2 supersede, 1 keep-both, 1 remove-link
• 3 duplicate clusters: 8 memories → 3
• 2 unscoped memories to re-scope
• 1 stale context memory to review
• 0 stale conversations

Approve all? Or review individually?

Wait for explicit approval. Do NOT execute anything until user approves.

Step 7: Execute approved changes

After user approves (all or specific items):

1. Run each approved operation:
   • memory_supersede for duplicates and outdated conflicts
   • memory_update for type/tag corrections
   • memory_delete only for worthless entries (user explicitly approved)
   • memory_link for relationship corrections
2. Report what was done, one line per action
3. Run memory_health() again, report improvement:

"Done. 142 → 134 memories. 4 → 0 conflicts. 3 duplicate clusters resolved."

Safety rules

These are hard rules. No exceptions.

• Never auto-delete. Every deletion requires user approval.
• Never auto-modify. All changes proposed first, executed after approval.
• Skip < 24h old memories. Too fresh to judge quality.
• Supersede over delete. Preserves history via links. Delete only for genuinely worthless entries.
• No new memories. This skill cleans. /rekal-save stores.
• No bulk approve without listing. Always show what will change before asking for approval.

Large databases (500+ memories)

Do NOT audit everything in one pass. Prioritize in this order:

1. Conflicts — highest impact on search quality
2. Topic clusters with count >= 5 — most duplicate accumulation
3. Recent timeline (last 30 days) — freshest quality issues

After completing priority items, ask:

"Covered conflicts and top duplicate clusters. Want me to continue with a deeper audit?"

Source: janbjorge/rekal — distributed by TomeVault.