redact

name	redact
description	Remove all agent-memory references to a session, mechanism, or identifier
argument-hint	<identifier> [--dry-run] [--keep <file>] [--keep-all-detail]

Redact — Purge Agent Memory References

Remove all references to a given identifier (session ID, mechanism name, etc.) from agent memory files. Project files (sessions/, tier0-computation/, constraint-map) are preserved — only .claude/agent-memory/ is touched.

Usage

/redact 29Ac                          # Remove all memory references to "29Ac"
/redact 29Ac --dry-run                # Show what would be removed, don't edit
/redact "rolling quintessence"        # Remove references to a mechanism name
/redact 21b --keep constraint-map.md  # Skip constraint-map (already default)
/redact 29Ac --keep-all-detail        # Only touch MEMORY.md files, skip detail files

Parse Arguments

1. Extract <identifier> — the first argument (or quoted multi-word string).
2. Extract flags:
   • --dry-run: Preview only, no edits.
   • --keep <file>: Skip files whose basename matches <file>. Multiple --keep flags allowed.
   • --keep-all-detail: Only edit files named MEMORY.md. Skip all other .md files in agent-memory dirs.
3. If no identifier provided, show the usage block above and stop.

Implementation Steps

Step 1: Scan

Use the Grep tool to search for <identifier> (case-insensitive) across all files matching the glob .claude/agent-memory/**/*.md.

Collect all matching files and their line numbers. Use output_mode: "content" with line numbers enabled to see exact matches.

If zero matches found, report:

=== REDACT: "<identifier>" ===
Scanned: .claude/agent-memory/**/*.md
Matches: 0

Nothing to redact.

Stop here.

Step 2: Filter files

Remove from the match list:

• Any file whose basename is constraint-map.md (default preserve — constraint-map is reference data that survives redaction).
• Any file matching a --keep <file> argument (basename comparison).
• If --keep-all-detail: any file NOT named MEMORY.md.

Track kept files separately — they will appear in the final report as "KEPT".

Step 3: Classify removals

For each remaining file with matches, read the full file using the Read tool. Classify each match into one of:

a) Section removal: If the match falls inside a section (delimited by ## or ### headers) where ALL or MOST content is about the identifier, mark the entire section (header through to next header or EOF) for removal.

b) Line removal: If the match is a single line within a larger section that covers multiple topics, mark just that line for removal. This includes:

• Bullet points (- ...)
• Numbered list items (1. ...)
• Standalone sentences

c) Table row removal: If the match is inside a markdown table (| ... |), mark just that row for removal. Never remove the header row or separator row.

d) Multi-reference line — FLAG: If a line contains BOTH the identifier AND other substantive, unrelated content (e.g., mentions both 29Ac and 29Ab with distinct information about each), do NOT remove it. Flag it for manual review in the report.

Key rules:

• Look for related file path references too: sessions/session-XX/, sXX_*.py, sXX_*.npz, sXX_*.png, sXX_gate_verdicts.txt where XX maps to the identifier.
• When removing a section, also remove any blank line immediately after the section that would create a double-blank-line.
• When removing a list item, if it was the last item under a header, also remove the now-empty header.
• Never remove the file's top-level # title header.

Step 4: Preview

Build a summary table showing what will be changed:

=== REDACT: "<identifier>" ===
Scanned: N agent memory files across M agents
Matches: X references in Y files

  agent/MEMORY.md ................. remove §Section Name (K lines)
  agent/detail-file.md ............ remove 2 lines from §Other Section
  agent2/MEMORY.md ................ remove 1 table row (1 line)
  agent3/MEMORY.md ................ FLAG: line N contains mixed refs (manual review)
  constraint-map.md ............... KEPT (default preserve)

If --dry-run, display this summary and stop. Do not make any edits.

Step 5: Execute edits

For each file with removals:

1. Use the Edit tool to remove the identified sections/lines.
2. Work from bottom-to-top within each file (highest line numbers first) to avoid offset drift.
3. After all removals, clean up artifacts:
   • Collapse any resulting triple-or-more blank lines down to a single blank line.
   • Remove any orphaned section headers (headers with no content before the next header).

Step 6: Verify

Use the Grep tool to re-scan for <identifier> across all .claude/agent-memory/**/*.md files.

• References in kept files (constraint-map, --keep targets) are expected — report them as "remaining (kept)".
• References in edited files indicate incomplete removal — report them as "WARNING: residual reference".
• Flagged multi-reference lines are expected — report them as "remaining (flagged for manual review)".

Step 7: Report

Show the final summary:

Before: NNNN lines across N files
After:  NNNN lines across N files
Removed: NN lines (N.N%)
Remaining: N references in kept/flagged files

If any files were flagged for manual review, list them with line numbers:

Manual review needed:
  agent/MEMORY.md:42 — mentions both "29Ac" and "29Ab"

Safety Rules

1. NEVER touch files outside .claude/agent-memory/ — sessions/, tier0-computation/, researchers/, tools/ are untouchable.
2. NEVER remove a line containing other valuable data — flag it for manual review instead.
3. NEVER remove constraint-map.md content by default — it's reference data. Only remove if the user explicitly passes --no-preserve-constraint-map (not a standard flag — they'd need to explicitly ask).
4. Section-level granularity preferred — removing a whole section is cleaner than leaving an orphaned header with one remaining bullet.
5. Table integrity — never remove a table's header row or separator row. Only remove data rows.
6. Preserve file structure — every file should still have its # title after redaction. If redaction would empty a file entirely, leave the title and a single line: <!-- Redacted: all content referenced <identifier> -->.

Error Handling

• If .claude/agent-memory/ directory doesn't exist: report error and stop.
• If identifier is empty: show usage and stop.
• If identifier is too short (1-2 chars): warn that this may match broadly and ask for confirmation before proceeding.
• If --keep file doesn't exist in the match list: silently ignore (it just means there were no matches in that file).