Assistants that remember who they are.
Persistent persona, on-disk memory, governed permissions. Wake them with cron, mentions, or just by talking. The same identity across Claude Code, Codex, and Gemini.
--resume. No active tool calls.Summarize PR reviews from this week (todo)
- Description: pull from acme-api, group by reviewer.
- Priority: normal · Created Mon at 10:25 PM
acme-api. I'll mark the task doing and update the team channel when I have a draft.Identity that survives restarts
Each assistant lives in {root}/assistants/{id}/. Open the folder, read the prompts, edit the memory. There's no black box.
Persona, constitution, persistence
Every assistant has a name, a persona, an org-and-personal constitution, and a memory dump — stored on disk, regenerated on change, byte-identical AGENT.md and CLAUDE.md.
Three layers of context, not three full prompts
Stable context (AGENT.md / CLAUDE.md) loads once. The Glueprint overlay handles primary-channel rules and safety. Only the per-wake user message changes per turn.
Long-term memory + journal
Typed memory files (user / project / feedback / reference) plus a daily journal. The assistant rebuilds an index on every write; you edit it in the Memory tab.
Wake on cron, channel, or chat
Routines fire on schedule. @mentions in channels wake the assistant. Talking to it in the chat resumes the same Claude Code session via --resume.
Three harnesses, one identity
Claude Code, Codex, and Gemini all read the same on-disk identity. Codex and Gemini inline the overlay into the first user message; Claude consumes it as an append-system-prompt-file.
Governed by your org, not by Glueprint
Permission tiers, tool filters, budgets, and quiet hours are server-authoritative and mirrored to disk so the assistant works the same way everywhere.
Bring Glueprint to your workflow.
Download once, run on your hardware, scale up to teams when you're ready. No credit card.