Memory

Every agent keeps a persistent memory — facts about you, your preferences, ongoing work, and the corrections you've given. It carries across conversations, so an agent you've worked with for a month knows things a fresh one wouldn't.

How it works

Memory is agent-scoped: your researcher's memory is separate from your writer's. An agent updates its own memory as it works — when you tell it something worth keeping ("always write dates as ISO"), it records the fact and applies it from then on.

A short summary of each agent's memory is woven into everything they do, and they can look deeper when a task calls for it.

Where it lives

On disk, in plain Markdown, under the agent's own folder:

~/.openacme/agents/<agent>/memory/

You can read it, edit it, or delete entries — it's your data, in files. The agent picks up your edits on its next conversation.

Shaping what's remembered

• Tell them directly. "Remember that our launch is June 30" lands in memory.
• Correct them once. Corrections are the highest-value memories — agents store the why, not just the what.
• Size limit. Each agent has a memory budget (default ~4,000 characters for the always-loaded summary) so memory stays sharp instead of becoming a junk drawer. Older, less-used detail moves to topic files that are loaded on demand.

Memory vs. shared context

Memory is per-agent and learned over time. For things every agent should know from day one — what your company does, naming conventions, who's who — use the shared AGENTS.md context file instead (web: Settings → Context).