Context Engineering for AI Agents: Lessons (opens in new tab)

+1 for using the file system as memory. You don't need a memory backend for that, or MCP, RAG, nothing. Just the basic read/write fs tools any agent has at its disposal.

Use an .agent/ directory. Tell it to keep a task list, a plan, findings, wishlists, future prompt ideas in there. I use it so much I have ".agent/" in my .gitignore template already. Tell it to use meaningful file names and to set the status of the memory using uppercase suffixes (and category as prefixes or subdirs), and you end up with...

    .agent/task-cleanup-code.DONE
    .agent/memory-function-style.md
    .agent/prompt-remove-dead-code.md
    .agent/symbols-map.toml
    ...

The agent will scan that directory every once in a while and look for information before making a decision. Granted, this is not a vectorial or graph search, it's a ripgrep regex probably, but the model can extrapolate and search many times for related words if needed.

Interesting that many of the same things are best practices for code repos. Don't let them bloat but don't refactor too often. Don't delete bad commits, but push a rollback PR instead. Don't get too repetitive or you might see imaginary patterns.

I can't think of an analog for the "don't delete tools" one though.

On fixed price plans, the provider has the incentive to set up caching since it improves their margins. On metered plans, the burden is on the user to pay attention to caching availability, and people are just beginning to appreciate its importance.