Memory — Active Write & Consolidate

The project's memory lives in two locations that must stay identical for safe (public) files (idempotent):

1. ~/.claude/projects/<escaped-cwd>/memory/ — harness, loaded by Claude Code, written first
2. .memory/ — canonical, git-tracked, written second — safe files only

Resolve the harness path: ~/.claude/projects/<cwd-with-/→->/memory/ (probe both space-handling forms: /→-+ →- and /→-+space-kept).

Private files (user preferences, credentials, PII) live in harness only. They must never appear as files or index lines under .memory/.

Active Write

When you encounter a decision, preference, lesson, or anything worth remembering, write it immediately — do not wait for /memory:consolidate.

Before writing

1. Search existing memories first (Grep/Read harness *.md by theme keywords). If one already covers the topic, edit that file instead of creating a near-duplicate.
2. Refuse pure operation logs: if you cannot state a durable Why and a reusable How to apply, do not write a memory. Put timelines in git/commits/docs, not memory.
3. Prefer one decision per file. Two independent decisions → two files; the same decision learned twice → one file.

Privacy check

.memory/ is a public GitHub repo. Private = secrets/PII/credentials or user preferences / personal workflow habits. Safe technical content → both locations. Private → harness only; harness index line ends with (harness only).

How to write

1. Write file to harness .../memory/<filename>.md
2. If safe → write identical file to .memory/<filename>.md. If private → do not write to .memory/; if .memory/<filename>.md already exists, delete it.
3. Update indexes:
   • Harness MEMORY.md: include every file (private lines marked (harness only))
   • .memory/MEMORY.md: safe lines only — never copy (harness only) lines into the public index

File naming: <type>_<kebab-slug>.md (type: feedback, project, reference)

Format:

---
name: <kebab-slug>
description: <one-line hook distinguishing this from similar files>
type: feedback | project | reference
---

<the fact>

**Why:** <why this decision exists>

**How to apply:** <actionable rules>

**Related:** [[other-memory]] [[another-memory]]

Memory is decision log, not operation log

Every memory file answers two questions only:

• Why — why this decision or rule exists
• How to apply — what to do next time

Remove all operation history (version numbers, dates-as-timeline, "first X then Y"). That lives in git log. Keep only the durable rationale and actionable rules.

MEMORY.md index format

Each line: one concise sentence, no version numbers, no date ranges, no timeline descriptions. One line per surviving file. Prefer ≤50 lines for scanability; if more files exist, keep every entry and group by theme — never drop entries to hit 50.

Good: feedback_git_commit_hook_needed.md — git PreToolUse hook intercepts git add/commit, redirects to /git:commit; allows chain + GIT_SKILL_FALLBACK=1 escape Bad: feedback_git_commit_hook_needed.md — git PreToolUse hook intercepts git add/commit; v0.5.3 command position anchoring + two exceptions + 26 regression tests

Red lines

• Never drop [[name]] cross-links when rewriting — preserve all from the original unless the target is intentionally deleted and the reference is removed in the same pass
• Never delete a file referenced by [[name]] in another memory file unless the reference is also removed
• rm only under the harness memory directory or .memory/ — never elsewhere
• Never publish private content to .memory/ (file body or index line)

Consolidate (/memory:consolidate)

User-invoked only. No auto-consolidation. Default failure mode of a weak run is cosmetic tidy while leaving thematic redundancy and factually dead notes. This procedure is fail-closed against that.

Work order: harness first, then sync to .memory/.

CRITICAL: Mutation freeze until planning artifacts exist

Do not Write, Edit, or rm any memory file until all three artifacts exist in this conversation:

1. Inventory — complete name list of harness *.md and .memory/*.md (including both MEMORY.md)
2. Cluster map — every non-index file appears in exactly one theme cluster
3. Staleness table — every non-index file has a rubric verdict (step 4)

After ground-truth probes (step 5), also hold a ground-truth table with tool-observed paths (path → found|missing|updated) before applying claim fixes. Fabricating these tables without Read/Grep/find/ls is a failed run.

CRITICAL: Machine validator (cannot self-attest past this)

Write inventory / cluster / staleness / report to temp files, then run:

# Pre-mutation (lift freeze only if exit 0)
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/validate-consolidate.py" \
  --inventory /tmp/mem-inventory.txt \
  --cluster /tmp/mem-cluster.txt \
  --staleness /tmp/mem-staleness.txt \
  --check=cluster,staleness

# Post-sync before claiming done (exit 0 required)
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/validate-consolidate.py" \
  --inventory /tmp/mem-inventory.txt \
  --cluster /tmp/mem-cluster.txt \
  --staleness /tmp/mem-staleness.txt \
  --report /tmp/mem-report.md \
  --harness "<harness-memory-dir>" \
  --public "<repo>/.memory"

Exit 1 → fix artifacts / privacy and re-run. Exit 0 is required to lift the mutation freeze (pre) and to report consolidate complete (post). Do not claim G2/G3/G4/privacy by prose alone.

CRITICAL exit gates (must all pass before reporting done)

Do not claim consolidate complete unless every gate below is true:

If any gate fails mid-run, continue until it passes — do not stop at "normalized frontmatter + rebuilt index".

1. Read every file

Read every *.md in both harness memory and .memory/, including MEMORY.md. Detect drift (name sets and content hashes / word-level diffs). List harness-only private files. Emit the inventory.

2. Normalize shape plan (do not stop here; mutations still frozen)

Plan only until artifacts in the freeze section exist:

• Relative dates → absolute YYYY-MM-DD
• Frontmatter: name, description, type only (strip node_type, originSessionId, modified, nested metadata)
• description specific enough to distinguish similar files
• Ensure Why / How to apply sections exist; if missing, derive them or mark for prune as non-decision

Normalization alone is not consolidation.

3. CRITICAL: Theme-cluster before merge

Group every non-index file into theme clusters (e.g. deploy, billing, review pipeline, naming). Use overlapping keywords, shared [[links]], and near-duplicate descriptions.

Output a cluster map (keep it for the report):

cluster: <theme>
  - file-a.md
  - file-b.md
  merge-default: yes|no — <one line>

Default bias: 2+ files in one theme → merge into one decision log, unless each holds a distinct durable decision that would become muddled if combined.

4. CRITICAL: Staleness rubric (practical expiry ≠ calendar expiry)

Score every non-index file. Calendar age is only one signal — a note can be days old and still SUPERSEDED or OPS-ONLY.

Verdicts: CONTRADICTED | SUPERSEDED | SUBSUMED | OPS-ONLY | ONE-SHOT | DORMANT | KEEP

Read references/staleness-examples.md for the full table, actions, and worked examples. Protect feedback_* preferences: incident dates do not make them OPS-ONLY.

5. CRITICAL: Ground-truth verify

For each project_* (and any reference_* that asserts repo-local facts):

1. Locate the claimed path, flag, API, architecture fact, or workflow in the current working tree
2. Record: VERIFIED | UPDATED | PRUNED | N/A (no repo) | UNVERIFIABLE (state why) plus tool-observed path → found|missing
3. Never leave a known-false claim in place after consolidate

Use Grep, Read, and allowed Bash (find/ls/stat/diff/…) against the project — not arbitrary shell. Do not trust memory text over the tree.

feedback_* about user/process preferences: verify consistency with other feedback files and current plugin/skill code when they name a mechanism; do not invent user-preference changes. Prefer KEEP on preference files unless contradicted by the user's later explicit preference.

6. Deduplicate and merge within clusters

Mutation freeze lifts only after inventory + cluster map + staleness table exist and pre-mutation validate-consolidate.py --check=cluster,staleness exits 0. For claim edits, also hold the ground-truth table first.

• Merge duplicates; keep the most detailed durable rules
• Collapse near-duplicates that differ only in incident detail
• Preserve all unique [[name]] targets; rewrite links to survivors after renames/merges
• After merge, survivor must still be decision-log shaped (Why + How), not a concatenated scrapbook

7. Rewrite for concision

• Why — root cause or decision rationale (1–3 short paragraphs max)
• How to apply — actionable bullets
• Related — [[name]] cross-links
• Strip version theater, step timelines, and "we tried X then Y"

8. CRITICAL: Independent adversarial pass

Run a second pass with clean context when any of: starting count ≥ 8 (excluding MEMORY.md); any cluster has 2+ files; ≥3 merges/deletes; user mentioned redundancy/stale memory; or you are unsure about keep-separate.

How: launch Agent (subagent_type general-purpose or Explore; fresh context). Pass inventory, cluster map, and paths only — not your keep-separate justifications. Ask it to propose merges, flag CONTRADICTED/SUPERSEDED/OPS-ONLY, list near-duplicate descriptions, and default to merge or prune when uncertain. Apply accepted findings; reject with one-line reasons in the report.

Zero findings despite multi-file clusters → re-check the largest cluster yourself. Skip only when count ≤ 5, all clusters size 1, no mutations, and ground-truth found nothing stale — state the skip. Never skip because the first pass "felt thorough".

9. Rebuild indexes (split public/private)

Rewrite two indexes:

1. Harness MEMORY.md — one line per surviving harness file; private lines end with (harness only)
2. .memory/MEMORY.md — one line per safe survivor only; strip every (harness only) line

No version/date theater. Group by theme if helpful. Keep every required entry (no hard 50-line drop).

10. Sync to .memory/ (privacy fail-closed)

For each file in harness memory:

1. If safe → write identical content to .memory/
2. If private → do not write to .memory/; delete .memory/<same name> if it exists (stale public copy)
3. Delete *.md in .memory/ that are not among current safe harness files
4. Never delete harness private files by "syncing absence" from a partial .memory/ read

After sync: public safe sets match; no private bodies or private index lines remain under .memory/.

rm targets only paths under harness memory or .memory/.

11. Report (required sections)

## Consolidate report
- Inventory: harness N, .memory/ M, drift: …
- Clusters: (map covering every non-index file)
- Staleness: verdict counts
- Ground truth: rows with path → found|missing + VERIFIED/UPDATED/PRUNED counts
- Merged: (old files → survivor) × reasons
- Pruned: (file × rubric verdict × reason)
- Kept-separate: (cluster × justification)
- Privacy: private files kept harness-only; public scrub actions
- Adversarial pass: ran|skipped — findings applied/rejected
- Index rebuilt: harness yes|no; .memory yes|no (safe-only)
- Validator: pre exit=…; post exit=… (paste PASSED/FAILED summary)
- Residual risks: anything still fuzzy
- Gates: G1–G8 checklist

Write the report body to a temp file and run the post-sync validator (full checks). Exit non-zero → do not claim done. If nothing changed after a full gated run, still show inventory, cluster map, ground-truth, and validator output.

Anti-patterns (do not do these)

• "Normalized frontmatter and rebuilt MEMORY.md" as the whole job while thematic near-duplicates remain
• Pruning only by "older than 3/6 months" while leaving last week's superseded architecture notes
• Creating a new memory file without searching for an existing theme file to edit
• Concatenating three incident writeups and calling it a merge (that is still an ops log)
• Skipping adversarial review because the first pass "felt thorough"
• Reporting done without G1–G8, without pre-mutation inventory/cluster/staleness, or without validator exit 0
• Copying the same MEMORY.md (including (harness only) lines) into both locations
• Leaving a private file that was once copied into .memory/ after a later privacy classification
• Self-attesting cluster/path/privacy in prose while skipping validate-consolidate.py