construct gives agents a small, durable memory surface for facts that should carry across turns, sessions, or future work in the same project. Memory is plain Markdown so it stays readable, editable, and easy to audit.
Memory is shared across all construct harness types in the same scope, so Codex, Claude Code, smith, and other agents can build on the same durable context.
Memory is intentionally separate from transcripts:
- Transcripts preserve what happened in a session.
- Memory preserves what is worth reusing later.
agentd_context exposes two memory files when available:
| Scope | Use for |
|---|---|
| Global memory | Cross-project preferences, standing workflows, and durable operating conventions. |
| Project memory | Project-specific architecture, workflows, decisions, commands, glossary, and pitfalls. |
Use the narrowest scope that will be useful later. A repo-specific workflow belongs in project memory; a preference the user repeats across repositories belongs in global memory.