tudoumashu/ai-memory-skillpack

What it solves

It prevents "context pollution" in AI agents by replacing unbounded, append-only handoff notes and changelogs with a durable, size-capped memory system. This ensures that agents do not waste thousands of tokens on outdated information between sessions and prevents the accidental leaking of secrets into project memory.

How it works

The system implements a tiered memory architecture (HOT, WARM, and COLD) for repositories, where only essential snapshots (HOT) are read every session, while detailed architecture and runbooks (WARM) are accessed via search, and history (COLD) is only read upon explicit request. It uses a "constitution" (SYSTEM.md) to enforce byte budgets and attribution, and an agent-native installer that allows AI agents like Codex or Claude Code to set up the system autonomously. To ensure reliability, it includes a verifier with executable assertions and SHA256 checksums to guarantee pack integrity.

Who it’s for

Developers using AI agents (specifically Codex CLI and Claude Code) who want a structured, maintainable way for their agents to remember project context across different sessions without bloating the context window.

Highlights

• Tiered Memory: Categorizes information into HOT, WARM, and COLD tiers to optimize token usage.
• Agent-Native Installation: Designed to be installed by the AI agent itself via an executable INSTALL.md.
• Fail-Closed Security: Includes a constitution that forbids storing secrets and a verifier that uses exit codes to prove installation success.
• Integrity Protection: Uses receipt-based hash comparisons to ensure upgrades do not silently overwrite user edits.