[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7ADQH5WGGEXSSWWQG85HXJ6":3,"skill-same:01M7ADQH5WGGEXSSWWQG85HXJ6":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":32,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":34,"install":35,"resources":48},"01M7ADQH5WGGEXSSWWQG85HXJ6","memory-hygiene",null,"Use when deciding whether to use cross-session agent memory (MEMORY.md, auto-memory), when auto-memory seems stale or bloated, when preferences leak across unrelated tasks, or when choosing between stateless and memory-equipped operation. Covers memory accumulation, temporal contamination, reasoning drift, and user-authored vs auto-saved preferences.","Leostruka\u002Fdevin-bundle","Leostruka","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F81370685?v=4","01M7AD8MAK63VDNAWRDRRRW04N","leandro ostruka","01M7AD8MBS65YDS1KMMV7YPM07",0,"MIT","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fmemory-hygiene\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: memory-hygiene\ndescription: Use when deciding whether to use cross-session agent memory (MEMORY.md, auto-memory), when auto-memory seems stale or bloated, when preferences leak across unrelated tasks, or when choosing between stateless and memory-equipped operation. Covers memory accumulation, temporal contamination, reasoning drift, and user-authored vs auto-saved preferences.\ntriggers: [user, model]\n---\n# Memory Hygiene\n\nCross-session agent memory is a **contract with the user**, not a black box. Naive accumulation degrades reliability; managed memory helps. The default is **explicit user-authored preferences**, not auto-saved memory.\n\n## Decision framework: stateless vs managed vs naive\n\n| Mode | When | Evidence |\n|---|---|---|\n| **Stateless** (no cross-session memory) | Default for coding agents. Preferences live in `.devin\u002Fglobal_rules.md` \u002F `.devin\u002Frules\u002F*.md` \u002F skills \u002F repo docs that the user controls. | Most predictable baseline; no drift, no contamination. |\n| **Managed memory** (selective add+delete) | Long-horizon tasks where recalling past executions genuinely helps (multi-session chat, long doc analysis). User reviews and prunes periodically. | +10% absolute vs naive growth (arXiv:2505.16067). MemGPT outperforms on multi-session chat (arXiv:2310.08560). |\n| **Naive auto-memory** (append-only, no pruning) | Never. | 16-20pp reliability loss (arXiv:2605.07313). Temporal contamination rises with exposure (arXiv:2605.17830). Reasoning drift even when answers look plausible (arXiv:2607.02374). |\n\n**Rule of thumb:** if a preference matters, the user should write it in `.devin\u002Fglobal_rules.md` or a skill. If past executions matter, use managed memory with selective addition and deletion - never append-only growth.\n\n## When to use this skill\n\n- Deciding whether to enable auto-memory \u002F MEMORY.md in an agent.\n- Auto-memory file is growing large or old.\n- Preferences from one task are leaking into unrelated tasks.\n- Choosing between stateless and memory-equipped operation for a new project.\n- Reviewing whether an agent's behavior has drifted from its baseline.\n\n## When NOT to use\n\n- Within-session context window management - use `context-window-hygiene` instead.\n- Offloading large context to files - use `context-folding`.\n- Selective cross-session notes - use `project-memory` for user-approved, plain-text, auditable memory.\n- The task is short and self-contained - stateless is correct, no decision needed.\n\n## Rules\n\n1. **Prefer explicit.** User-authored preferences (`.devin\u002Fglobal_rules.md`, `.devin\u002Frules\u002F*.md`, skills, repo docs) over agent-auto-saved memory. The user controls the steering.\n2. **Never naive growth.** If auto-memory is used, it must have selective addition AND deletion. Append-only is the anti-pattern.\n3. **Audit periodically.** If auto-memory exists, review it manually for stale, large, or drifting entries. Prune aggressively.\n4. **Isolate preferences by task.** A preference about feature X should not influence work on feature Y. Auto-memory that globally injects all preferences into every session causes contamination (arXiv:2605.17830).\n5. **Preference following \u003C10% at 10 turns** (arXiv:2502.09597). Don't rely on the agent to infer and persist preferences - it won't, reliably. Write them down.\n6. **Memory is a contract.** If the agent saves something to memory, the user must be able to view, modify, and delete it (arXiv:2404.15269, CIPHER). Opaque auto-memory violates the contract.\n\n## Anti-patterns\n\n- **Append-only MEMORY.md.** Grows forever, never pruned. Maximizes contamination and drift.\n- **Auto-saving every preference.** \"The user said they like tabs → save to memory → inject into every future session.\" Causes a preference about one feature to leak everywhere.\n- **Trusting auto-inferred preferences.** Preference following accuracy is \u003C10% at 10 turns zero-shot (arXiv:2502.09597). Auto-inferred preferences are unreliable.\n- **\"Delete all memory.\"** The opposite extreme. Managed memory helps (+10% vs naive, arXiv:2505.16067; MemGPT, arXiv:2310.08560). The problem is naive accumulation, not memory itself.\n- **Opaque memory.** Auto-memory the user cannot inspect or edit violates the contract.\n\n## Academic basis\n\n- **Memory accumulation degrades reliability.** Shao et al. 2026, \"When Stored Evidence Stops Being Usable\" (arXiv:2605.07313): HippoRAG loses 16-20 percentage points in budget-compliant reliability as irrelevant sessions accumulate, even with task evidence held fixed.\n- **Temporal memory contamination.** Al-Tawaha et al. 2026, \"Remembering More, Risking More\" (arXiv:2605.17830): memory-induced violation rates show a robust upward trend with exposure length; effect driven by accumulated content, not encounter order. Tested on 8 memory architectures + Claw-like agents.\n- **Error propagation + misaligned replay.** Xiong et al. 2025, \"How Memory Management Impacts LLM Agents\" (arXiv:2505.16067): naive memory growth causes error propagation (inaccuracies compound) and misaligned experience replay (outdated\u002Firrelevant experiences harm current tasks). Selective add+delete yields +10% absolute vs naive growth.\n- **Reasoning drift from memory.** Fang et al. 2026, DRIFTLENS (arXiv:2607.02374): user-attribute memory induces medium-to-large reasoning drift above the pragmatic-noise floor, even when final answers remain fluent and plausible. Partly mitigable (GRPO\u002FDPO), not eliminable.\n- **Preference following is unreliable.** Zhao et al. 2025, PrefEval (arXiv:2502.09597, ICLR 2025 oral): preference following accuracy \u003C10% at 10 turns (~3k tokens) zero-shot across most models. Even with prompting+retrieval, deteriorates in long context.\n- **User-editable preferences improve alignment.** Lin et al. 2024, CIPHER\u002FPRELUDE (arXiv:2404.15269): learning descriptive preferences from user edits, with user view\u002Fmodify access, outperforms direct edit retrieval and context-agnostic preferences.\n- **Managed memory helps.** Packer et al. 2023, MemGPT (arXiv:2310.08560): OS-inspired memory hierarchy outperforms fixed-context LLMs on multi-session chat and document analysis. Wu et al. 2024, LongMemEval (arXiv:2410.10813): optimized memory designs (session decomposition, fact-augmented keys) improve recall and QA; 30% accuracy drop without optimization shows memory is hard but valuable.\n- **Excessive retrieval harms agentic tasks.** \"Harness the Memory\" (arXiv:2608.15008): retrieving more entries helps QA but hurts sequential decision-making - attention shifts from action-critical context to retrieved blocks. Substrate routing is necessary.\n\n## Source\n\nDistilled from AI-coding workflow guidance on memory hygiene. Claims verified against primary sources - the thesis (A: accumulation pollutes, B: stateless is predictable, C: user-authored > auto-saved) is supported; the conclusion (D: delete all memory) is refuted by MemGPT and managed-memory evidence. The correct prescription is \"prefer explicit, allow managed, ban naive.\"\n","# Memory Hygiene\n\nCross-session agent memory is a **contract with the user**, not a black box. Naive accumulation degrades reliability; managed memory helps. The default is **explicit user-authored preferences**, not auto-saved memory.\n\n## Decision framework: stateless vs managed vs naive\n\n| Mode | When | Evidence |\n|---|---|---|\n| **Stateless** (no cross-session memory) | Default for coding agents. Preferences live in `.devin\u002Fglobal_rules.md` \u002F `.devin\u002Frules\u002F*.md` \u002F skills \u002F repo docs that the user controls. | Most predictable baseline; no drift, no contamination. |\n| **Managed memory** (selective add+delete) | Long-horizon tasks where recalling past executions genuinely helps (multi-session chat, long doc analysis). User reviews and prunes periodically. | +10% absolute vs naive growth (arXiv:2505.16067). MemGPT outperforms on multi-session chat (arXiv:2310.08560). |\n| **Naive auto-memory** (append-only, no pruning) | Never. | 16-20pp reliability loss (arXiv:2605.07313). Temporal contamination rises with exposure (arXiv:2605.17830). Reasoning drift even when answers look plausible (arXiv:2607.02374). |\n\n**Rule of thumb:** if a preference matters, the user should write it in `.devin\u002Fglobal_rules.md` or a skill. If past executions matter, use managed memory with selective addition and deletion - never append-only growth.\n\n## When to use this skill\n\n- Deciding whether to enable auto-memory \u002F MEMORY.md in an agent.\n- Auto-memory file is growing large or old.\n- Preferences from one task are leaking into unrelated tasks.\n- Choosing between stateless and memory-equipped operation for a new project.\n- Reviewing whether an agent's behavior has drifted from its baseline.\n\n## When NOT to use\n\n- Within-session context window management - use `context-window-hygiene` instead.\n- Offloading large context to files - use `context-folding`.\n- Selective cross-session notes - use `project-memory` for user-approved, plain-text, auditable memory.\n- The task is short and self-contained - stateless is correct, no decision needed.\n\n## Rules\n\n1. **Prefer explicit.** User-authored preferences (`.devin\u002Fglobal_rules.md`, `.devin\u002Frules\u002F*.md`, skills, repo docs) over agent-auto-saved memory. The user controls the steering.\n2. **Never naive growth.** If auto-memory is used, it must have selective addition AND deletion. Append-only is the anti-pattern.\n3. **Audit periodically.** If auto-memory exists, review it manually for stale, large, or drifting entries. Prune aggressively.\n4. **Isolate preferences by task.** A preference about feature X should not influence work on feature Y. Auto-memory that globally injects all preferences into every session causes contamination (arXiv:2605.17830).\n5. **Preference following \u003C10% at 10 turns** (arXiv:2502.09597). Don't rely on the agent to infer and persist preferences - it won't, reliably. Write them down.\n6. **Memory is a contract.** If the agent saves something to memory, the user must be able to view, modify, and delete it (arXiv:2404.15269, CIPHER). Opaque auto-memory violates the contract.\n\n## Anti-patterns\n\n- **Append-only MEMORY.md.** Grows forever, never pruned. Maximizes contamination and drift.\n- **Auto-saving every preference.** \"The user said they like tabs → save to memory → inject into every future session.\" Causes a preference about one feature to leak everywhere.\n- **Trusting auto-inferred preferences.** Preference following accuracy is \u003C10% at 10 turns zero-shot (arXiv:2502.09597). Auto-inferred preferences are unreliable.\n- **\"Delete all memory.\"** The opposite extreme. Managed memory helps (+10% vs naive, arXiv:2505.16067; MemGPT, arXiv:2310.08560). The problem is naive accumulation, not memory itself.\n- **Opaque memory.** Auto-memory the user cannot inspect or edit violates the contract.\n\n## Academic basis\n\n- **Memory accumulation degrades reliability.** Shao et al. 2026, \"When Stored Evidence Stops Being Usable\" (arXiv:2605.07313): HippoRAG loses 16-20 percentage points in budget-compliant reliability as irrelevant sessions accumulate, even with task evidence held fixed.\n- **Temporal memory contamination.** Al-Tawaha et al. 2026, \"Remembering More, Risking More\" (arXiv:2605.17830): memory-induced violation rates show a robust upward trend with exposure length; effect driven by accumulated content, not encounter order. Tested on 8 memory architectures + Claw-like agents.\n- **Error propagation + misaligned replay.** Xiong et al. 2025, \"How Memory Management Impacts LLM Agents\" (arXiv:2505.16067): naive memory growth causes error propagation (inaccuracies compound) and misaligned experience replay (outdated\u002Firrelevant experiences harm current tasks). Selective add+delete yields +10% absolute vs naive growth.\n- **Reasoning drift from memory.** Fang et al. 2026, DRIFTLENS (arXiv:2607.02374): user-attribute memory induces medium-to-large reasoning drift above the pragmatic-noise floor, even when final answers remain fluent and plausible. Partly mitigable (GRPO\u002FDPO), not eliminable.\n- **Preference following is unreliable.** Zhao et al. 2025, PrefEval (arXiv:2502.09597, ICLR 2025 oral): preference following accuracy \u003C10% at 10 turns (~3k tokens) zero-shot across most models. Even with prompting+retrieval, deteriorates in long context.\n- **User-editable preferences improve alignment.** Lin et al. 2024, CIPHER\u002FPRELUDE (arXiv:2404.15269): learning descriptive preferences from user edits, with user view\u002Fmodify access, outperforms direct edit retrieval and context-agnostic preferences.\n- **Managed memory helps.** Packer et al. 2023, MemGPT (arXiv:2310.08560): OS-inspired memory hierarchy outperforms fixed-context LLMs on multi-session chat and document analysis. Wu et al. 2024, LongMemEval (arXiv:2410.10813): optimized memory designs (session decomposition, fact-augmented keys) improve recall and QA; 30% accuracy drop without optimization shows memory is hard but valuable.\n- **Excessive retrieval harms agentic tasks.** \"Harness the Memory\" (arXiv:2608.15008): retrieving more entries helps QA but hurts sequential decision-making - attention shifts from action-critical context to retrieved blocks. Substrate routing is necessary.\n\n## Source\n\nDistilled from AI-coding workflow guidance on memory hygiene. Claims verified against primary sources - the thesis (A: accumulation pollutes, B: stateless is predictable, C: user-authored > auto-saved) is supported; the conclusion (D: delete all memory) is refuted by MemGPT and managed-memory evidence. The correct prescription is \"prefer explicit, allow managed, ban naive.\"\n",{"name":5,"triggers":24,"description":7},[25,26],"user","model","main","02b95ddb62959a00b3c31442c2046aca0d19563c","skills\u002Fmemory-hygiene",1,"https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle","https:\u002F\u002Fleostruka.github.io\u002Fdevin-bundle\u002F","https:\u002F\u002Fgithub.com\u002FLeostruka","2026-11-15 05:19:58",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add Leostruka\u002Fdevin-bundle --skill memory-hygiene",{"tool":41,"command":42},"bunx","bunx skills add Leostruka\u002Fdevin-bundle --skill memory-hygiene",{"tool":44,"command":45},"pnpm","pnpm dlx skills add Leostruka\u002Fdevin-bundle --skill memory-hygiene","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fmemory-hygiene\u002FSKILL.md",true,[],[50,57,64,71,78,85],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M7E0K3MG2XC4FSRV0MK6SFR1","montador-de-ideias","Use when assembling, expanding, or elevating a raw project idea or initial prompt into an ultra-detailed, first-principles dossier and compiled master prompt with journalistic OSINT market research, Tech Lead and Staff Architect justifications, atomic and modular hardening\u002Foptimization, and high-impact contact surfaces.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fmontador-de-ideias\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M7E0J7JJWZ8QMK01RZ3B3C9Z","git-github-adaptive-engineering","Use when any task touches git history, GitHub issues\u002FPRs\u002Freleases\u002FCI, repo governance, or branch and merge decisions. Runs the Fase Zero situational-awareness battery, picks one of five delivery flows via the adaptive decision engine, and enforces commit\u002FPR information-density standards. Orchestrates the gh, git-workflows, github-actions, pr, and debug\u002Ffinish skills.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fgit-github-adaptive-engineering\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M7E0GJFRPFEVQQMXEM58EWDM","adaptive-agenda-calendar-routine-engine","Use when planning or executing work across hours, days, weeks, or sprints. Turns the calendar into a strict time-blocked agenda (HH:MM-HH:MM blocks carrying a routine profile, expected inputs, and a handoff output artifact), runs the 5-stage cognitive flow (Radar and Hunt -> Triage plus Steve Jobs conception -> workbench manifest -> >=32-cycle loop -> Close plus Open-Minded Scrum), and unfolds macro calendar milestones into micro agenda blocks through extensions\u002Fgcal\u002Fwrapper.py.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Fskills\u002Fadaptive-agenda-calendar-routine-engine\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M7AE2K9FTY0BVS48N0A9JFKF","demo","Use when demonstrating a skill in a fixture project.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Fdemo\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M7AE2G3DP38PTV3B1N7MW5TS","duplicate-skill","Use when testing duplicate skill name detection.","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Fbeta\u002FSKILL.md",[],[],{"skillNo":86,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M7AE2DBR8ZHT6T875MJ392GK","https:\u002F\u002Fgithub.com\u002FLeostruka\u002Fdevin-bundle\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fdevin-config\u002Fproject\u002F.devin\u002Fskills\u002Falpha\u002FSKILL.md",[],[]]