Use when the user asks to "remember project context"; manages the SEO/GEO memory lifecycle — hot-cache, active work, archive tiers, and privacy cleanup. Not for content or domain scoring — use the auditors. 项目记忆/跨会话
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
memory-management
description
Use when the user asks to "remember project context"; manages the SEO/GEO memory lifecycle — hot-cache, active work, archive tiers, and privacy cleanup. Not for content or domain scoring — use the auditors. 项目记忆/跨会话
This skill implements a three-tier memory system (HOT/WARM/COLD) for SEO and GEO projects. HOT memory (80 lines max) loads automatically every session via the SessionStart hook. WARM memory loads on demand per skill. COLD memory is archived data queried only when explicitly requested. The skill manages the full lifecycle: capture, promote, demote, and archive.
What This Skill Does
Manages a three-tier memory lifecycle (HOT/WARM/COLD) with automatic promotion, demotion, and archival. Also maintains open-loop tracking and cross-skill aggregation.
Quick Start
Start with one of these prompts. Finish with a hot-cache update plan and a handoff summary using the repository format in Skill Contract.
Initialize Memory Structure
Set up SEO memory for [project name]
Initialize memory structure for a new [industry] website optimization project
Update After Analysis
Update memory after ranking check for [keyword group]
Refresh hot cache with latest competitor analysis findings
Query Stored Context
What are our hero keywords?
Show me the last ranking update date for [keyword category]
Look up our primary competitors and their domain authority
Promotion and Demotion
Promote [keyword] to hot cache
Archive stale data that hasn't been referenced in 30+ days
Glossary Management
Add [term] to project glossary: [definition]
What does [internal jargon] mean in this project?
Skill Contract
Expected output: a memory update plan, hot-cache changes, and a short handoff summary.
Reads: current campaign facts, new findings from other skills, approved decisions, and the shared State Model.
Writes: updates to memory/hot-cache.md, memory/open-loops.md, memory/decisions.md, and related memory/ folders. Manages WARM-to-COLD archival in memory/archive/. Auditor handoff archiving (v7.1.0+): when triggered by a direct user request or an auditor's explicit "Save these results?" yes-response, append a structured block to . The Stop hook never initiates memory writes. See for the exact archive block format and rules.
Promotes: durable strategy, blockers, terminology, entity candidates, and major deltas. Applies the temperature lifecycle by observable rules only: promote to HOT on an explicit user/skill request ("promote X" / pin); demote and archive by the file's last_updated date. Reference-frequency counters are NOT tracked by any hook, so the lifecycle never depends on them — see State Model.
Done when: the requested lifecycle action (capture/promote/demote/archive/query/purge) is applied, memory/hot-cache.md is within the 80-line / 25KB limit, and the affected memory paths are reported back to the user.
Primary next skill: use the Next Best Skill below when the project memory baseline is ready for active work.
This skill's behavior is reinforced by the library's claude-hook.sh hooks. What the hook actually does (do not document behavior it lacks):
SessionStart (fires on startup, resume, clear, compact): injects a sanitized excerpt of memory/hot-cache.md, and when memory/open-loops.md has tracked items, appends a one-line pointer to review them for staleness. It does not compute dates or a "Quick Status" — surfacing which open loops are stale is the agent's job once pointed at the file.
PostToolUse: warns when memory/hot-cache.md exceeds 80 lines / 25KB; enforces the auditor artifact-gate on memory/audits/*.md writes; offers an optional quality check after user-facing content edits.
Stop: a no-op (exits without output). CLAUDE.md's "allow-only Stop check" is just this no-op; the hook never initiates memory writes.
Data Sources
With tools: auto-populate from ~~SEO tool, ~~analytics, ~~search console. Without tools: ask user for keywords, competitors, metrics, campaigns, and terminology. See CONNECTORS.md.
Decision Gates
Stop and ask the user when:
A purge (Art 17 / CCPA) is requested — present the matched files and the redaction-vs-delete choice, and only act on confirmed matches. Never auto-delete memory.
A memory/decisions.md entry needed to answer a query has approved_by: skill_inferred or a missing field — surface it as ADVISORY and confirm before treating it as authoritative.
A referenced term is not found in any memory layer — ask for clarification rather than guessing.
Continue silently (never stop for):
Routine promotion/demotion that follows the temperature lifecycle rules.
Hot-cache trimming suggestions when over the 80-line / 25KB limit (recommend, don't block).
Missing optional tool data when auto-populating — record what is available and proceed.
Instructions
When a user requests SEO memory management:
1. Initialize Memory Structure
For new projects, create the directory structure defined in the State Model. Key directories: memory/ (decisions, open-loops, glossary, entities, research, content, audits, monitoring).
When a user references something unclear, follow this lookup sequence:
Step 1: Check memory/hot-cache.md (hot cache)
Is it in active keywords?
Is it in primary competitors?
Is it in current priorities or campaigns?
Step 2: Check memory/glossary.md
Is it defined as project terminology?
Is it a custom segment or shorthand?
Step 3: Check Cold Storage
Search memory/archive/ first for dated YYYY-MM-DD- archived files.
If the archive points to a source category, follow that trail back to memory/research/, memory/audits/, or memory/monitoring/.
Treat COLD findings as historical unless refreshed by the current session.
Step 4: Ask User
If not found in any layer, ask for clarification
Log the new term in glossary if it's project-specific
Decision provenance (v8.0.1+): when loading memory/decisions.md, verify each entry has approved_by: user. Entries with approved_by: skill_inferred or missing field are treated as ADVISORY — surface to user before using as authoritative. Auditor-class skills (content-quality-auditor, domain-authority-auditor) MUST ignore non-user-approved decisions when determining verdict. See skill-contract.md §Promotion Rules.
Example lookup: User asks "Update rankings for our hero KWs" → Step 1 finds "Hero Keywords (Priority 1)" in hot-cache → extract the keyword list → run the ranking check → update memory/hot-cache.md and memory/monitoring/rank-history/YYYY-MM-DD-ranks.csv.
3. Promotion & Demotion Logic
Reference: See Promotion & Demotion Rules for detailed promotion/demotion triggers (keywords, competitors, metrics, campaigns) and the action procedures for each.
Reference: See Update Triggers & Integration for the complete update procedures after ranking checks, competitor analyses, audits, and reports; monthly/quarterly archive routines; and integration points with all 8 connected skills (keyword-research, rank-tracker, competitor-analysis, content-gap-analysis, seo-content-writer, content-quality-auditor, domain-authority-auditor).
5. Memory Hygiene Checks
When invoked for review or cleanup:
Line count check: Count lines in memory/hot-cache.md. If >80, list oldest entries for archival.
Byte check: If hot-cache exceeds 25KB, warn and recommend trimming long entries.
Staleness scan: List memory files whose frontmatter last_updated date (or file mtime) is older than 30 days; recommend archival for files older than 90 days. Age is computable from disk — reference-frequency is not tracked, so never gate on "unreferenced".
Frontmatter audit: Check that all memory files (except hot-cache.md) have name, description, and type in their frontmatter. Report any missing fields.
6. Save Results
Ask "Save these results for future sessions?" — if yes, write YYYY-MM-DD-<topic>.md to memory/. Add veto issues to memory/hot-cache.md only from auditor handoff or explicit user approval.
GDPR / Privacy Compliance
memory/ may store third-party personal data — entity names, founder bios, LinkedIn profiles, author/journalist names surfaced by entity-optimizer or research skills. Under GDPR Art 4(1) (applies to processing of personal data of EU/EEA/UK residents regardless of where the controller is located), these qualify as "personal data". The user is the data controller. Non-EU users without EU/EEA/UK data subjects may still face analogous obligations under CCPA/CPRA (California), PIPEDA (Canada), LGPD (Brazil), or other national regimes. Not legal advice.
Retention policy
WARM files: archive to memory/archive/ after 90 days unreferenced (default lifecycle)
COLD archive: never auto-deleted, but eligible for Art 17 erasure requests
All files: user MUST honor Art 17 requests from data subjects (individuals named in memory)
Greps all files under memory/ (including memory/archive/) for the entity name, slug, or domain — grep -rF "<entity-name>" memory/ — and presents matches for confirmation.
On confirmation, deletes or anonymizes the matched lines/files across the working tree: memory/hot-cache.md, WARM notes, COLD/archive files, memory/entities/<slug>.md, memory/entities/candidates.md, audit aggregates, and open loops.
Appends a dated, subject-free entry to memory/audits/gdpr-purges.md per the GDPR Purge Log Template — required fields date, redacted_label, legal_basis, action, scope, working_tree_only: true — so there is a human-readable record of the request.
Honest limitation — this edits the working tree only. If memory/ is under version control (it usually is, in the user's project repo), the subject still exists in git history. Verify with git log -S"<entity-name>" -- memory/; true erasure from history requires git filter-repo / git filter-branch and is the user's responsibility — it is out of this skill's scope. Do not represent a working-tree redaction as a complete, audit-grade erasure. There is no salted-fingerprint or reingest-blocking mechanism: nothing in the hooks consults a tombstone before writing, so any such claim would be false.
Lawful basis reminder
Before writing a third-party person to memory/entities/, the user must have one lawful basis per GDPR Art 6 (where GDPR applies — see scope note above): consent, legitimate_interest, contract, or equivalent. Advisory — this skill does not enforce, and does not substitute for legal review.
Reference Materials
Examples — Worked examples, advanced features, practical limitations, and the auditor handoff archive block format & rules
