Memory Index Compact

Memory Index Compact

Purpose / When to Activate

Activate when:

• A registry file (index.md, index-*.md, archive index) breaches its file-size budget (typically 40k chars). The post-commit hook surfaces this.
• An index row has accreted substance that duplicates the target file body (Tier 2 cycle trails, commit SHAs, §-numbered substance dumps, validation ceremony, drift-heal forensic prose). Detection : any row > 200 chars in a registry table is suspect.
• Scheduled hygiene pass during memory-index-sync or post-DEC-activation cleanup.

Distinct from other memory skills:

• memory-compact — compacts memory CONTENT (sessions, domains) into summaries. NOT for index files.
• memory-archive-superseded — migrates superseded DEC rows to archive. Adjacent but separate.
• memory-index-sync — heals drift between filesystem and registry (missing rows). Detects bloat ; compaction is THIS skill's job.

Core principle — index = pointer

Every row in a registry must answer one question : « Does an agent reading this row for relevance assessment need this word? »

• Keep : file path, ID, category, level, status, ≤30-word topic essence, relation to other entities (amends, supersedes, related).
• Dégage : Tier 2 cycle trails, finding counts (F-C1, REFINE 2C+7H+4M+4L), commit SHAs, validation ceremony ("Fred ratified post Tier 2 + Compliance 11/11 PASS"), §-numbered substance dumps duplicating DEC body, drift-heal forensic prose, self-referential pointers without payload.

Substance lives in:

• DEC body (decisions/DEC-XX.md) — for substantive design rationale
• _log.md — for forensic governance trail (cycle history, status flips, drift heals)
• git log — for commit history
• artefact files (contexts/artefacts/research/*.md) — for deep dives

Index rows duplicating any of these are anti-pattern.

Process

1. Identify target file. Read the registry. Confirm it is an index (rows = pointers, not content). Skip if file is substantive content with embedded tables (those follow memory-compact flow).

2. Measure baseline. Record current size in chars. Note the budget (typically 40k for memory index files — check hooks/post-commit.d/03-memory-size-budget.sh BUDGETS array). Calculate target reduction.

3. Triage rows. For each row > 200 chars, classify content per the « keep / dégage » heuristic:

   • Keep : metadata (path, ID, status, date) + ≤30-word topic + key relations (amends/supersedes).
   • Dégage : everything else, mapping each cut to its source-of-truth file (DEC body / _log.md / git / artefact).

4. Compact the row. Rewrite in the target form:

   • Active Files table row: | path | category | ID | YYYY-MM-DD (status) |
   • Decision registry row: | DEC-XX | domain | level | ≤30-word topic. Amends/supersedes X. | status YYYY-MM-DD |
   • Changelog row: | YYYY-MM-DD | ≤15-word event. Detail → _log.md / DEC-XX. |
   • External artefacts row: | path | ≤15-word topic | referenced-by IDs |

5. Compact narrative sections. Same principle for non-table prose:

   • Navigation / TL;DR blocks: keep only actionable pointers, no philosophy.
   • Section preambles ("How agents read this section") : ≤2 sentences max.
   • Cascade summaries (paragraphs duplicating registry data) : delete outright.
   • Frontmatter updated_at with embedded change history : reduce to date only.

6. Verify no information loss. Before any cut, confirm the dropped content lives in another source-of-truth file. If a row's content does NOT live elsewhere, the substance must move TO the appropriate target file FIRST (then cut from index).

7. Update budget tracking. If the post-commit hook flagged the breach, verify new size is ≤ budget. Re-run the hook locally if available.

Outputs

• Updated index file(s) — pointer-only rows, all substance preserved in source-of-truth files
• No new files (this skill compacts in place ; for extract-to-sibling pattern, see precedent §7.4 in strategy/strategic-frame.md)
• No deletions of source-of-truth files

Quality Checks

• Every row is ≤200 chars unless an explicit relation chain justifies more (e.g. multi-DEC amends)
• No row contains : commit SHAs, Tier 2 finding counts, §-numbered substance dumps, validation ceremony prose
• All metadata fields preserved : path, ID, category, level, status, date
• Target file size ≤ budget (per hook BUDGETS array)
• memory-index-lint passes post-compaction (no broken row symmetry vs filesystem)
• Spot-check : pick 3 random compacted rows ; can an agent decide relevance from them ? If yes, PASS.

Non-Goals

This skill must NOT:

• Move or delete source-of-truth files (decisions/DEC-XX.md, _log.md, artefacts)
• Extract content into sibling files (that is the §7.4 file-extraction pattern, separate concern)
• Archive superseded rows (use memory-archive-superseded)
• Compact memory CONTENT — sessions, domains, patterns body (use memory-compact)
• Add new metadata to rows (drift-heal new fields belongs in memory-index-sync)
• Modify DEC bodies, _log.md, or any source-of-truth file as part of compaction (substance must already live there before the cut — verify in step 6)

Relationship with other skills

file-size budget breach → memory-index-compact (this skill)
filesystem-vs-registry drift → memory-index-sync
DEC superseded → memory-archive-superseded
content categories oversized → memory-compact
validation invariants → memory-index-lint

If unsure which skill applies, the failure mode of each is well-bounded :

• memory-index-compact cuts text but never moves files. Reverse via git revert.
• memory-compact archives ephemeral content. Reverse via archive restore.
• memory-archive-superseded migrates rows + sets archived_to: frontmatter. Reverse via git revert.

Pointer principle is invariant. Even if other skills create or amend rows, those new rows MUST be in pointer-only form per this skill's heuristic. Substance-duplication in any row is a defect.