Menu
Use when user wants to view, modify, or reset memory system configuration, or check token usage across memory tiers
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Use when user asks about project status, wants to rename/move a project, clean up stale data, or mentions orphaned folders in ~/.claude/
Use when user wants to load a project's long-term and short-term memory mid-session — typically when running in light mode and Claude lacks project-specific context
Use when processing pending session transcripts into daily summaries — first session of the day, on manual request, or on schedule
Use when user wants to audit long-term memory for stale, incorrect, outdated, or duplicate entries, or correct memory entries based on new information
Use when user asks about past work, decisions, discussions, or topics that might have relevant history beyond loaded context
Use when user wants to explicitly preserve something permanently in long-term memory's Pinned section
| name | settings |
| description | Use when user wants to view, modify, or reset memory system configuration, or check token usage across memory tiers |
| user-invokable | true |
View, modify, and analyze the memory system configuration.
/settings or /settings view — Show current settings/settings usage — Show token usage breakdown/settings set <path> <value> — Modify a setting/settings reset [path] — Reset to defaults (all or specific setting)/settings or /settings view:~/.claude/memory/settings.json/settings usage:Run the token usage script and display results:
uv run --no-project $HOME/.claude/scripts/token_usage.py
Display as a usage table with ~Tokens, Limit, % Used, and status indicator (within limit vs over limit).
/settings set <path> <value>:projectShortTerm.workingDays)~/.claude/memory/settings.jsonNote: Short-term tokenLimit and totalTokenBudget are calculated automatically from workingDays (formula: workingDays * 750).
/settings reset [path]:Reset to defaults from _defaults section of settings.json. Omit path to reset all.
| Path | Type | Range | Default | Notes |
|---|---|---|---|---|
mode | string | full/light | full | light skips project LTM/STM and global STM (recall + global LTM only) |
globalLongTerm.tokenLimit | int | 1000-50000 | 3,000 | Fixed limit |
globalShortTerm.workingDays | int | 1-30 | 2 | Also updates tokenLimit |
projectLongTerm.tokenLimit | int | 1000-50000 | 3,000 | Fixed limit |
projectShortTerm.workingDays | int | 1-30 | 5 | Also updates tokenLimit |
synthesis.intervalHours | float | 0.25-24 | 1 | In-script rate-limit floor (hours). Timer cadence is fixed at install time from the default, so values below the default only affect the floor; values above lengthen the effective cadence. |
synthesis.model | string | haiku/sonnet/opus | sonnet | Model for synthesis subagent |
synthesis.minSessionMessages | int | 0-100 | 5 | Skip sessions with fewer messages during synthesis |
decay.ageDays | int | 7-365 | 30 | Archive learnings older than this |
decay.projectWorkingDays | int | 5-100 | 20 | Archive project entries after N project work days |
decay.archiveRetentionDays | int | 30-730 | 365 | Purge archived items older than this |
| Component | Default | Formula |
|---|---|---|
| Global long-term | 3,000 | fixed |
| Global short-term | 1,500 | workingDays * 750 |
| Project long-term | 3,000 | fixed |
| Project short-term | 3,750 | workingDays * 750 |
| Total | 11,250 | sum of above |
Estimation: 1 token ~ 4 characters. 750 tokens/day based on ~400-600 observed after scope filtering.
Mode: full loads all five sections (recall, global LTM, project LTM, project STM, global STM). light loads only recall and global LTM — useful when Claude Code's native project-scoped memory covers project context and you want the custom system to handle only cross-project memory plus session continuity.
Working days: Both tiers use "working days" — days that have matching tagged content ([global/*] or [project-name/*]). Days without matching content are skipped.
Git subdirectories: Subdirectories of a git repo automatically map to the repo root unless gitignored — no configuration required. Gitignored paths (e.g. projects/) stay as separate projects.
Synthesis: Runs via launchd/systemd timer (deferred). Manual /synthesize always runs in foreground.
Decay: Archives entries older than ageDays from decay-eligible sections (Key Actions/Decisions/Learnings/Lessons). ## Pinned section, auto-pinned sections, and entries without dates are protected. Archive at ~/.claude/memory/.decay-archive.md.
~/.claude/memory/settings.json — defaults from DEFAULT_SETTINGS in memory_utils.py if missing.