Systems on Glass is an AI-automation side project of mine. Every episode is generated by Claude (Anthropic's AI, working through Claude Code in my own video studio): the script, the collage and 3D scenes, the edit, and the narration in an AI clone of my voice. I choose the topics, bring the stories from my own work, and review each episode before it goes up. Not affiliated with Anthropic.
Halfway through making episode 8, the agent that makes this series ran out of context. Claude Code summarized the conversation, threw the rest away and kept going, with my voice settings, my colours and my rules intact. None of those lived in its memory. They lived in files.
The short version
- A model remembers nothing between calls. The context window is its RAM (episode 4): fast, small and wiped when the session ends. So everything worth keeping has to be written down and read back.
- Four kinds of memory. Princeton's CoALA paper splits an agent's memory like ours: working memory (right now), and three that last: semantic (facts), episodic (what happened) and procedural (how to do things).
- Facts, every time. Claude Code loads
CLAUDE.mdfrom your folder and every folder above it at launch. Write it concrete enough to verify: "Use 2-space indentation", not "Format code properly". - Your corrections are the memory. My studio's file has a section called "What the owner has taught us". I pointed out a repeated thumbnail photo once; now every episode gets a new image. Correct it once, write it down, and it stays corrected.
- Notes it keeps, know-how it opens. Agents can keep their own notes (Anthropic's Claude playing Pokémon tracked "1,234 steps" and "8 levels toward the target of 10"); long workflows live in skills, opened just in time. The catch: memory is context too, so keep each file under 200 lines.
compaction keeps a summary + the 5 most recently accessed files (Anthropic) · 4 kinds of memory (CoALA, 2023)
The examples are from my own studio: its CLAUDE.md, its contributing rules ("one video at a time" came from me asking why two renders were crawling) and its log of which photo opened each episode.
Further reading
- How Claude remembers your project — the Claude Code docs: where memory files live, when they load, how to write them.
- Cognitive Architectures for Language Agents — Sumers, Yao, Narasimhan and Griffiths (2023): the map of working, semantic, episodic and procedural memory.
- Effective context engineering for AI agents — Anthropic on compaction, structured note-taking and just-in-time context.
- The memory tool — a memory directory Claude checks before starting a task.
Next episode: Retries, backoff and jitter. Previously: Queues. All episodes →