🧠 Intent‑Scoped Memory — reference architecture for OpenClaw Discord bots
I wrote a concrete memory architecture/spec for OpenClaw Discord bots that scopes memory by intent rather than global recall.
It’s designed for private Discord servers and trusted collaborators, and treats:
• Discord channels as ephemeral attention (nothing remembered by default)
• Discord threads as isolated, durable memory
• 1:1 Discord DMs as shared context
It also introduces study mode: an explicit mode where a thread is treated as “we are studying this”, so the bot carefully ingests and remembers external material — including docs, PDFs, links, videos, and online courses — instead of skimming or guessing.
Optionally, the bot can run in an ambient mode where users don’t need to @mention it — it participates naturally when appropriate, without changing memory boundaries.
This is a reference architecture (not a plugin or SDK), intended to help Discord bot builders avoid context bleed and unpredictable memory behavior as systems grow.
GitHub: https://github.com/Threadm-ind/intent-scoped-memory
Happy to hear thoughts or edge cases people have run into 👋