Agent skill

recall-maintain

Pressure-based memory cleanup. Demotes, archives, or pins entries when working memory exceeds budget.

Stars 192
Forks 12

Install this agent skill to your Project

npx add-skill https://github.com/davegoldblatt/total-recall/tree/main/skills/recall-maintain

SKILL.md

Run maintenance on the memory system. Primary function: pressure-based demotion when working memory exceeds the word budget. Secondary: stale entry verification, contradiction checks, open loop review.

What To Do

0. Preconditions

Check for IDs first. Scan all managed files (CLAUDE.local.md, memory/registers/.md, memory/archive/**/.md) for single-line list item entries (lines starting with - , not inside code blocks, not placeholders).

  • If ANY managed entry lacks an ID (no ^tr[0-9a-f]{10} at end of line): refuse to run.

    Cannot run maintain: [N] entries are missing IDs.
    Run /recall-init-ids first to tag all entries.
    
  • If ANY duplicate IDs are found across entries: refuse to run.

    Cannot run maintain: duplicate ID found.
    ID: ^tr8f2a1c3d7e
      1. CLAUDE.local.md line 15
      2. registers/preferences.md line 8
    Resolve manually before running maintain.
    

If preconditions pass, proceed.

1. Load and Reconcile

Parse entries from all managed files. For each entry, extract:

  • The entry text (everything between - and ^tr...)
  • The ID (the ^tr[0-9a-f]{10} token)
  • The file path and line number
  • The tier (derived from file location):
    • CLAUDE.local.md -> working
    • memory/registers/* -> register
    • memory/archive/* -> archive

Load metadata from memory/.recall/metadata.json:

  • If file doesn't exist, create it as empty {}
  • For each inline ID missing from metadata, create a metadata entry with created_at=now, last_reviewed_at=now, pinned=false, snoozed_until=null, status=active, tier from file
  • For each metadata entry, recompute tier from file location (file location is authoritative - overwrite metadata tier if it differs)

2. Compute Working Memory Pressure

Count words in working memory entries only (tier=working):

  • For each working entry, count whitespace-separated tokens in the entry text
  • Exclude the leading - and trailing ^tr... from word count
  • Sum all working entry word counts = working_words
  • Target = 1500 words

Report:

Working memory: [working_words] words (target: 1500)

3. Select Candidates

A. Pressure Candidates (only if working_words > 1500)

From working entries, exclude:

  • Entries with pinned: true in metadata
  • Entries with snoozed_until in the future

Score remaining working entries:

score = 1.0 * word_count(entry) + 0.1 * days_since(last_reviewed_at)

(If last_reviewed_at is missing, treat as 365 days.)

Sort by score descending. Take the top candidates until the sum of their word counts >= (working_words - 1500). This is the minimum set needed to bring working memory under budget.

B. Superseded Cleanup (always, regardless of pressure)

Find any entry across all tiers where metadata has status: "superseded" and tier is NOT archive. These should be archived.

4. Present Candidates

If no candidates (under budget and no superseded entries):

recall-maintain — all clear
━━━━━━━━━━━━━━━━━━━━━━━━━━

Working memory: [N] words (target: 1500) - under budget
No superseded entries pending cleanup.
No action needed.

Otherwise, present each candidate with its reason:

recall-maintain — [N] candidates
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Working memory: [working_words] words (target: 1500, over by [N])

Pressure candidates (need to free [N] words):

1. [CLAUDE.local.md] 23 words, last reviewed 45 days ago
   "Dave prefers concise error messages and no unnecessary logging"
   → [k]eep / [p]in / [s]nooze / [d]emote / [a]rchive / [m]ark superseded

2. [CLAUDE.local.md] 18 words, last reviewed 30 days ago
   "Current deployment target is AWS us-east-1 with fallback to us-west-2"
   → [k]eep / [p]in / [s]nooze / [d]emote / [a]rchive / [m]ark superseded

Superseded cleanup:

3. [registers/tech-stack.md] status: superseded
   "Using Node 16 in production"
   → [a]rchive (recommended) / [k]eep

Apply all / decide individually?

5. Execute Actions

For each candidate, apply the chosen action:

keep: Update last_reviewed_at to now in metadata. No file changes.

pin: Set pinned: true and update last_reviewed_at in metadata. No file changes. Entry will be excluded from future pressure candidates.

snooze: Set snoozed_until to now + 30 days (or user-specified duration) and update last_reviewed_at in metadata. No file changes. Entry will be excluded from pressure candidates until the snooze expires.

demote (working -> register):

  1. Remove the entry line from CLAUDE.local.md
  2. Append the entry line (with ID) to memory/registers/_inbox.md
    • Create _inbox.md with header if it doesn't exist:
      markdown
      # Inbox
      > Entries demoted from working memory. Review and file into appropriate registers.
      
  3. Update metadata tier to register and last_reviewed_at to now

archive (working or register -> archive):

  1. Remove the entry line from its current file
  2. Append the entry line (with ID) to memory/archive/ARCHIVE.md
    • Create ARCHIVE.md with header if it doesn't exist:
      markdown
      # Archive
      > Archived memory entries. Searchable but never auto-loaded.
      
  3. Update metadata: set status: "archived", tier: "archive", last_reviewed_at to now

mark_superseded: Set status: "superseded" and last_reviewed_at to now in metadata. No file changes. Entry will appear in superseded cleanup on next maintain run.

File Editing Rules

When moving entries between files:

  • Remove the ENTIRE line (including the - prefix and ^tr... suffix)
  • Do NOT reflow, reformat, or modify any other lines in the source file
  • Append the ENTIRE line to the destination file
  • Preserve exact formatting of the entry text

6. Write Metadata

After all actions are applied, write memory/.recall/metadata.json:

  • Sort keys alphabetically for readable diffs
  • Use 2-space indentation
  • Write deterministically (same input = same output)

7. Secondary Checks

After the pressure-based flow, run these additional checks (same as before):

Stale entries: Search registers for entries where metadata last_reviewed_at is older than 30 days. Present them:

Stale entries (not reviewed in 30+ days):
  1. [registers/tech-stack.md] Last reviewed: 2025-12-01
     "Using Postgres 15 in production" ^tra1b2c3d4e5
     → [v]erify (update last_reviewed_at) / [u]pdate / [a]rchive

Contradictions: Scan across tiers for conflicting claims on the same topic. Flag any found.

Open loop review: Check memory/registers/open-loops.md for items that are past due or resolved.

Daily log archival: If daily logs older than 30 days exist, suggest archiving to memory/archive/daily/.

8. Summary

recall-maintain — complete
━━━━━━━━━━━━━━━━━━━━━━━━━━

Working memory: [before] -> [after] words (target: 1500)

Actions taken:
  Kept (reviewed): [N]
  Pinned: [N]
  Snoozed: [N] (until [date])
  Demoted to _inbox.md: [N]
  Archived: [N]
  Marked superseded: [N]

Secondary checks:
  Stale entries verified: [N]
  Contradictions found: [N]
  Open loops reviewed: [N]
  Daily logs to archive: [N]

Metadata: [N] entries in memory/.recall/metadata.json

Expand your agent's capabilities with these related and highly-rated skills.

Didn't find tool you were looking for?

Be as detailed as possible for better results