Cog: Persistent Plain-Text Memory for Claude Code

Cog is a plain-text architecture that gives Claude Code persistent memory through Markdown files and Unix tools. It features a nightly consolidation pipeline that acts as 'REM sleep' to help the AI refine its rules and learn from daily interactions. The project is an open-source experiment aimed at making AI cognition transparent, traceable, and capable of long-term growth.

Key Points

• Cog enables persistent memory for Claude Code using plain-text Markdown files rather than a database.
• The architecture includes a nightly 'REM sleep' pipeline for data consolidation, pattern extraction, and rule evolution.
• Memory is organized into a three-tier system (desk, filing cabinet, and deep storage) to optimize context usage.
• The system is fully transparent and user-editable, with all AI-driven changes recorded in a git log.
• Cog serves as an experimental platform to study AI self-reflection and knowledge management.

Sentiment

The community is cautiously interested but largely skeptical. While the underlying problem of persistent AI memory resonates with many developers who share their own approaches, the consensus leans toward simpler solutions being sufficient for most coding agent workflows. The AI-generated writing style of the article draws pointed criticism, and several commenters see the project as one of too many competing memory standards without clear differentiation.

In Agreement

• The three-tier memory structure with nightly consolidation mirrors how human memory works, with separate immediate, long-term, and archived layers that reference contextual associations
• Plain-text, git-tracked memory provides genuine transparency—even if models eventually handle memory internally, beliefs should remain inspectable and editable by the user
• Memory should be purpose-driven and personalized rather than a generic bucket, since what needs to be remembered varies drastically by user type and task
• Confidence scoring and contradiction logging complement Cog's freshness-based approach, addressing the problem that not all stored information is equally reliable

Opposed

• Simple approaches like CLAUDE.md, README files, and chat history already solve this problem adequately without the complexity of a full memory architecture
• Explicit rules and skills are less noisy, easier to maintain, and more effective than memory systems for coding agent use cases
• This is a superficial workaround—real memory capabilities need to be built into models through ML research, not scaffolded via prompt engineering
• The article itself is AI-generated and painful to read, undermining credibility by exemplifying the low-effort content the community dislikes
• Loading extensive memory context degrades LLM output quality due to long-context hallucination issues; minimal task-focused context produces better results