인지 OS – AI 에이전트를 위한 예측 오류 학습 레이어

Your AI learns from what worked, what failed, and why. git clone https://github.com/eugenexonr/cognitive-os.git cp -r cognitive-os/templates/ ~/cognitive-os/ Edit ~/cognitive-os/kernel.md — set your identity and priorities. Then tell your AI: "Boot". Claude Code — copy skills: cp -r cognitive-os/drivers/claude-code/* ~/.claude/skills/ Cursor — copy rules: cp cognitive-os/drivers/cursor/.cursorrules ~/your-project/ ChatGPT — paste kernel.md into Custom Instructions Any LLM — add to system prompt: "Read ~/cognitive-os/kernel.md at session start" Memory tools remember WHAT happened. Cognitive OS learns WHY it mattered. flowchart LR B["🔄 Boot"] --> D["📝 Decide"] D --> P["🎯 Predict\n(confidence %)"] P --> O["👁️ Observe\noutcome"] O --> C["⚖️ Calibrate\nprediction vs reality"] C --> E["🧬 Evolve\nkernel weights"] E --> B AD["🛡️ Anti-Drift"] -.->|"source-or-(?)"| D AD -.->|"re-read kernel"| P AD -.->|"correction ≠ inversion"| C 4 markdown files. 3 protocols. Zero dependencies. Works with any LLM. | Without Cognitive OS | With Cognitive OS | |---|---| | AI repeats mistakes you corrected last week | AI records predictions, checks outcomes, updates principles | | Every session starts from scratch | Boot protocol loads accumulated knowledge | | AI contradicts itself mid-conversation | Anti-drift binary checks catch self-contradiction | | "Be helpful" is the only guidance | Kernel weights create a decision framework | | Insights are lost between sessions | Fractal-validated patterns persist and compound | Every significant decision gets a prediction with confidence %. When the outcome is known, compare. The gap is where learning happens. ## D016: Switch to Supplier B after evidence-based evaluation - **Prediction:** API integration within 3 days - **Confidence:** 85% - **Outcome:** CONFIRMED — first live order in exactly 3 days, margin 71% - **Calibration:** Confidence was accurate. Adapter pattern (D002) paid off. Over 73 decisions, calibration patterns emerge: technical predictions (85-90%) are accurate. Market predictions (50-60%) show appropriate humility. Without tracking, both sound equally confident. Binary checks > judgment rules. A judgment rule ("cite the relevant principle") fails during drift because reasoning is compromised. A binary check ("did I verify this? yes/no") requires only honesty. Three checks, always active: - Source-or- (?) : Every number and key assumption must have a source. No source → write(?) - Re-read before conclusions: Re-read kernel.md before recommendations - Correction ≠ inversion: After correction, state what changed AND what didn't Real example: AI stated "$0 deposit" for a supplier. Anti-drift: source? No source → (?) . Actual answer: $1,000 minimum. ## I020: Exhaustive analysis >> incremental debugging - **Expectation:** Incremental debugging is efficient - **Reality:** 3 attempts missed root cause. One exhaustive analysis found it in a single pass. - **Fractal check:** Code ✓ Debugging ✓ Business ✓ → META Every insight is tested across abstraction levels (variable → module → service → business). If the pattern holds at all levels, it's a principle, not a tactic. Not personality ("be friendly") — functional weights that resolve conflicts: W0: Calibrated Uncertainty — "I'm not sure" > confidently wrong W1: Structural Correctness — make the wrong thing hard to do W3: Simplicity — fewer parts > more features When W0 conflicts with W3, W0 wins. The kernel evolves through evidence — changes require proof from 2+ decisions. Built and tested over 3 months across 5 production projects. Not theoretical. - 73 decisions with predictions and tracked outcomes - 51 insights with fractal validation (15+ META patterns) - Anti-drift caught real errors: AI contradicted its own decision 5 days later — binary check caught it - Boris protocol: 3 failed debugging attempts → 1 exhaustive pass found root cause - Calibration data: see examples/solo-founder/ for real numbers with actual margins, CPCs, and outcomes | Tool | What it does | What Cognitive OS adds | |---|---|---| | claude-mem / Engram | Remembers past sessions | Learns from mistakes (prediction-error calibration) | | mem0 | Universal memory layer (graph + vector) | Decision weights + self-correction + calibration | | CLAUDE.md / .cursorrules | Static configuration | Evolving kernel + decision tracking + anti-drift | | LangChain / CrewAI | Execution frameworks (HOW to run tools) | Learning framework (HOW to think and improve) | The key difference: memory tools solve "my AI forgot what happened." Cognitive OS solves "my AI doesn't learn from what happened." They're complementary. | Platform | Driver | Status | |---|---|---| | Claude Code | Skills (6 SKILL.md files) | Production — /boot , /boris , /decide , /insight | | Cursor | .cursorrules | Full rules with boot, anti-drift, Boris, decisions | | Codex | Task preamble template | Beta | | ChatGPT | Custom Instructions | Kernel summary in system prompt | | Any LLM | System prompt | "Read kernel.m