Chain-of-Thought (CoT) prompting has significantly enhanced the reasoning capabilities of Large Language Models (LLMs). However, CoT can also produce fluent but logically flawed reasoning chains that make incorrect answers appear more reliable. In this work, we introduce REMIND (Reliability Enhancement via Metacognitive INtrospection and Debiasing), a framework that enables LLMs to learn from their own failures via episodic memory. Unlike standard error-correction methods that rely on external feedback or extensive fine-tuning, REMIND leverages a self-supervised introspection cycle. We first isolate epistemic failure—instances where CoT decreases model confidence despite yielding incorrect results. The model then performs structured introspection to diagnose the root cause and synthesizes a reusable natural language lesson. These lessons are stored in a vector database and retrieved during future inference steps to correct similar future errors. Experiments on eight diverse reasoning benchmarks spanning mathematical, scientific, multi-hop, truthfulness, and commonsense reasoning demonstrate that REMIND significantly improves reliability by effectively converting past CoT failures into actionable insights for unseen problems. Codes are available in https://github.com/newcodevelop/REMIND.