# Continuous Learning System > Claude Codeセッションを信頼度スコアリング付きの小さな学習済み行動である「インスティンクト」を通じて再利用可能な知識に変える高度な学習システム。 Parent: [LLM Models and APIs](https://llms-explorer.com/tree/llm-models-and-apis/) · 7 facets · 29 facts · page: https://llms-explorer.com/tree/continuous-learning-system/ ## Continuous Learning v2 - インスティンクトベースアーキテクチャ - Claude Codeセッションを信頼度スコアリング付きの小さな学習済み行動である「インスティンクト」を通じて再利用可能な知識に変える高度な学習システム。 — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#continuous-learning-v2---インスティンクトベースアーキテクチャ) ## インスティンクトモデル - インスティンクトは小さな学習済み行動です： — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#インスティンクトモデル) - アトミック - 1つのトリガー、1つのアクション — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#インスティンクトモデル) - 信頼度重み付け - 0.3 = 暫定的、0.9 = ほぼ確実 — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#インスティンクトモデル) - ドメインタグ付き - code-style、testing、git、debugging、workflowなど — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#インスティンクトモデル) - 証拠に基づく - それを作成した観察を追跡 — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#インスティンクトモデル) ## 1. 観察フックを有効化 - ~/.claude/settings.jsonに追加します。 — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - プラグインとしてインストールした場合（推奨）： — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - { "hooks": { "PreToolUse": [{ "matcher": "", "hooks": [{ "type": "command", "command": "~/.claude/skills/continuous-learning-v2/hooks/observe.sh" }] }], "PostToolUse": [{ "matcher": "", "hooks": [{ "type": "command", "command": "~/.claude/skills/continuous-learning-v2/hooks/observe.sh" }] }] } } — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - mkdir -p ~/.claude/homunculus/{instincts/{personal,inherited},evolved/{agents,skills,commands}} touch ~/.claude/homunculus/observations.jsonl — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - /instinct-status # 信頼度スコア付きの学習済みインスティンクトを表示 /evolve # 関連するインスティンクトをスキル/コマンドにクラスター化 /instinct-export # 共有のためにインスティンクトをエクスポート /instinct-import # 他の人からインスティンクトをインポート — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - { "version": "2.0", "observation": { "enabled": true, "store_path": "~/.claude/homunculus/observations.jsonl", "max_file_size_mb": 10, "archive_after_days": 7 }, "instincts": { "personal_path": "~/.claude/homunculus/instincts/personal/", "inherited_path": "~/.claude/homunculus/instincts/inherited/", "min_confidence": 0.3, "auto_approve_threshold": 0.7, "confidence_decay_rate": 0.05 }, "observer": { "enabled": true, "model": "haiku", "run_interval_minutes": 5, "patterns_to_detect": [ "user_corrections", "error_resolutions", "repeated_workflows", "tool_preferences" ] }, "evolution": { "cluster_threshold": 3, "evolved_path": "~/.claude/homunculus/evolved/" } } — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) - ~/.claude/homunculus/ ├── identity.json # プロフィール、技術レベル ├── observations.jsonl # 現在のセッション観察 ├── observations.archive/ # 処理済み観察 ├── instincts/ │ ├── personal/ # 自動学習されたインスティンクト │ └── inherited/ # 他の人からインポート └── evolved/ ├── agents/ # 生成された専門エージェント ├── skills/ # 生成されたスキル └── commands/ # 生成されたコマンド — [source](https://llms-explorer.com/sources/mdb-context-hub/continuous-learning-v2/#1-観察フックを有効化) ## Where this helps - Claude Code setups that repeat the same corrections across many sessions — a preferred git workflow, a recurring tool preference, a repeated debugging pattern — and would benefit from those corrections being captured automatically. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Teams wanting to share learned behaviors across people or machines, since instincts can be exported and imported rather than staying locked to one person's session history. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Situations where low-confidence, still-forming behavioral patterns should accumulate evidence over multiple observations before being trusted, rather than acting on a single occurrence. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Growing a set of related, individually small instincts into a proper reusable skill, agent, or command once enough of them cluster around the same workflow. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Project ideas - Enable the PreToolUse/PostToolUse observation hooks in settings.json and let the system accumulate observations for a week of normal work before reviewing what instincts it proposed via /instinct-status. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Tune the confidence-decay rate and auto-approve threshold to match your risk tolerance — a stricter auto-approve threshold trades slower automation for fewer false-positive learned behaviors. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Use /evolve to promote a cluster of related instincts sharing a domain tag into a generated skill, agent, or command, then review the generated artifact before trusting it in production. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Export your personal instincts with /instinct-export and import a colleague's with /instinct-import to bootstrap a shared team baseline instead of starting from zero on a new machine. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Antipatterns - Enabling the observation hooks without ever reviewing /instinct-status, letting low-confidence or wrong instincts accumulate unchecked instead of being pruned or corrected. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Setting the auto-approve threshold too low, letting the system act on instincts that are still statistically thin (few supporting observations) as if they were well-established. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Importing another person's or team's instincts wholesale without reviewing them for relevance, since a workflow preference tuned to one codebase or role can actively mislead in a different context. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Treating an evolved skill or command generated from clustered instincts as finished the moment it's created, instead of reviewing it the way you would any other newly authored skill. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Known issues - Instincts are atomic by design (one trigger, one action), so a genuinely conditional or multi-step behavior may not be representable as a single instinct without decomposition. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Confidence scores are derived from the observations that produced them, so a system in early use with few observations will have systematically lower-confidence instincts regardless of how correct the underlying behavior is. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - The observer component runs on a lightweight model on a fixed interval, which trades observation cost against how quickly a new pattern is detected and turned into a candidate instinct. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Because instincts and observations are stored locally under ~/.claude/homunculus/, moving to a new machine without an explicit export/import step means the accumulated learning does not automatically follow you. — [source](https://llms-explorer.com/tree/continuous-learning-system/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Context files - [Continuous Learning System](https://llms-explorer.com/downloads/sources/mdb-context-hub/continuous-learning-v2.md)