memeri

Starting a new chat

Status: ✅ Current · Last reviewed: 2026-06-18

Memeri works best when your AI starts from shared context, not a blank slate. Set it up once, then every chat is grounded. Two cases: starting something new, and continuing something that exists.

You never paste a long prompt. You add a tiny one-time instruction; the agent fetches Memeri's current rules itself (get_memeri_bootstrap). The rules can evolve — your setup never goes stale.

One-time setup (per agent)

  1. Connect the Memeri connector (see Connect).
  2. Add the one-time instruction to the agent's standing instructions:
    • ChatGPT → Settings → Personalization → Custom Instructions
    • Claude Code → ~/.claude/CLAUDE.md
    • Codex → AGENTS.md

The instruction is the same for all:

When I explicitly ask to use Memeri … first call get_memeri_bootstrap and follow the returned instructions. If I discuss a project but don't mention Memeri, ask before using Memeri tools.

New project — "Set this up in Memeri"

Say: "Set this up in Memeri: <one-line idea>."

The agent checks your existing projects, creates a new one (never writing into an existing project without asking), builds the first workstream + jobs, and seeds Project Understanding. A real workspace from one sentence.

Existing project — "Get oriented first"

Say: "Get oriented on <project> first, then …" — or just start; a set-up agent self-orients.

The agent calls get_project_context (format "brief") — which now folds in the Project Understanding digest (current focus, key decisions, constraints) plus recent activity — then scoped get_project_context for your task, and get_project_understanding for the full source-backed map when it needs depth. It inherits the same understanding every previous session built, so it never starts from zero.

Normal chat

Don't mention Memeri and it stays a normal chat. The agent only reaches for Memeri when you ask.

What goes in CLAUDE.md (Claude Code)

Two slots, two jobs:

  • Global — ~/.claude/CLAUDE.md (every session, everywhere). Put the conservative pointer above. It makes Claude Code Memeri-aware without assuming every repo is a Memeri project.

  • Per-project — ./CLAUDE.md in a repo that is tracked in Memeri. Add an always-orient line, because here Memeri is the workspace:

    This project is tracked in Memeri (slug: your-slug). At session start, call get_project_context({ project: "your-slug", format: "brief" }) (it includes the Understanding digest), then scoped get_project_context; call get_project_understanding for the full map, and follow Memeri's working conventions via get_workflow_playbook("work_in_project"). After a context compaction, call get_session_recovery_context() first.

That work_in_project playbook is the live discipline — narrate progress as job updates, keep the roadmap current, record decisions, close out jobs — so your AI's work flows into the project instead of happening invisibly. It's fetched live, so it stays current with the platform; your pasted pointer never goes stale.

New = "Set this up in Memeri." · Existing = "Get oriented first." · Otherwise = just chat.