ECC: Memory System for Claude Code That Survives Sessions

You spend two hours pairing with Claude Code, building out a feature with careful architectural decisions and established patterns. Close the terminal. Reopen it the next morning. Your AI assistant has no memory of any of it—like meeting a stranger who happens to have access to your codebase.

ECC fixes this context-window amnesia. It's a framework that coordinates planning, test-driven implementation, fresh-context review, verification, persistent memory, and reusable skills instead of starting from scratch every session. The hooks summarize and re-inject session context so Claude Code picks up where you left off.

How Memory Works

ECC extends Claude Code through persistent hooks that survive terminal restarts. When you close a session, the framework captures decisions, patterns, and context. When you reopen, those hooks feed that information back in—your AI assistant remembers what you were building and why.

This isn't a replacement for Claude Code. It's a layer on top, built on Anthropic's SKILL.md mechanism and distributed as skills, agents, hooks, rules, and MCP configurations. The system treats memory as infrastructure, not an afterthought.

The Six-Phase Workflow

ECC structures AI-assisted development into coordinated phases: planning, test-driven implementation, fresh-context review, verification, persistent memory storage, and skill creation. Instead of typing the same context-setting prompts every morning, you get a system that builds on previous work.

Each phase feeds the next. Planning decisions become reusable skills. Implementation patterns get captured in hooks. Verification results inform future sessions. The workflow replaces fragmented repetition with continuity.

Shipped In Production

The maintainer built Zenith.chat entirely with ECC-enhanced Claude Code workflows and won the Anthropic x Forum Ventures hackathon in September 2025. The framework ships real projects under deadline pressure. When your memory system survives a hackathon sprint, it works.

The Token-Usage Trade

A concern from the community: skill descriptions enter context and get cached, but a large catalog increases token usage. This is a conscious tradeoff. You're paying for memory—persistent context across sessions costs tokens. For quick one-off tasks, vanilla Claude Code makes more sense. For multi-day projects where starting fresh wastes time, the token cost buys workflow continuity.

If you accumulate hundreds of skills, you'll see higher usage. The question is whether that cost is worth not re-explaining your architecture every morning.

Active Maintenance

Issue #521 exposed a memory-exhaustion bug where the Continuous Learning v2 observer crashed because every tool call triggered analysis without throttling or batching. The system was learning too aggressively, overwhelming itself. The issue was identified, discussed, and addressed—the kind of response that matters when a tool becomes load-bearing in your workflow.

ECC's progression through versions 2.0, 2.1, and 2.2 shows a maintainer iterating based on real use. Version 2.2.3 shipped in October 2025, incorporating lessons from production deployments.

Auto-Generated Project Memory

The ECC GitHub App analyzes your repository history and agent configuration, then opens pull requests with project-specific skills, rules, hooks, manifests, and safety checks. Instead of manually teaching your AI assistant about your codebase conventions, the app bootstraps that memory from existing code patterns.

This automates the initial setup cost—the framework learns your project's idioms and generates the hooks to maintain that context across sessions.

Who Needs This

If you're using Claude Code for quick scripts or isolated tasks, the default experience works fine. If you're building something over days or weeks—where continuity matters and re-explaining context wastes momentum—ECC's persistent memory pays off. The framework targets developers who've felt the pain of context-window amnesia and want their AI assistant to remember yesterday's decisions.

Claude Code and similar tools changed how we write code. ECC extends them with the one thing they lack: a memory that survives closing the terminal.