Hacker News Highlights Cog, a Plain-Text Memory Architecture for Claude Code

Why this Show HN post stood out

On March 25, 2026, a Show HN post sent readers to Cog, a "cognitive architecture" for Claude Code built almost entirely from plain-text conventions. The linked project describes itself as a set of rules rather than an app: CLAUDE.md defines behavior, markdown files hold memory, and the filesystem becomes the interface that Claude can inspect with familiar Unix tools such as grep, find, and git diff. That framing matters because many persistent-memory experiments for agents add databases, vector stores, or custom runtimes. Cog deliberately does not.

How the system is structured

The project splits memory into three tiers. "Hot" memory stays small and is loaded every session for current priorities. "Warm" memory lives in domain folders such as personal or work and is loaded when related skills activate. "Glacier" archives older material behind an index so context stays retrievable without bloating every prompt. The project also relies on L0/L1/L2 retrieval hints, where files carry brief summaries that help Claude decide whether to open them fully.

What makes Cog more than a folder template is the workflow layer. The project ships commands such as /setup, /reflect, /housekeeping, /evolve, and /foresight. According to the documentation, these routines condense observations into patterns, archive stale material, rebuild indexes, and even adjust the memory architecture over time. The important nuance is that the model itself is not changing. The rules and memory files are. That keeps every change inspectable in git instead of hidden inside opaque agent state.

Why it matters

For teams already using Claude Code, Cog is interesting because it treats persistent agent behavior as infrastructure you can review like any other project asset. There is no separate service to operate and no hidden memory layer to debug. If the agent behaves unexpectedly, the project argues that you should be able to inspect the relevant markdown, permissions, and recent diffs directly.

• Plain-text first: memory files stay readable and editable with standard tools.
• Workflow over magic: scheduled reflection and housekeeping are explicit commands, not implicit background behavior.
• Auditability: the git log becomes part of the memory story.

The tradeoff is also obvious. A file-based memory system only helps if the conventions stay disciplined. Hot memory must remain short, ownership of facts must stay clear, and users still need to decide what deserves promotion into persistent state. Even so, Cog captures a direction many developer-agent workflows are moving toward: less hidden orchestration, more inspectable text.

Community source: Hacker News discussion. Original project: marciopuga/cog and project docs.