🧠Intent‑Scoped Memory
A people‑aligned memory architecture for Discord bots — designed for OpenClaw. Channels forget. Threads remember. 1:1 DMs share. Intent decides.
Intent‑Scoped Memory defines where memory is allowed to exist in a Discord bot — and just as importantly, where it is not — based on how people actually use Discord.
This architecture was designed and validated using OpenClaw as the primary reference implementation. It can be adapted to other Discord bot frameworks as a secondary use case.
The problem
Most Discord bots fail socially, not technically.
They:
- leak memory across conversations
- bias new answers with old context
- treat channels, threads, and DMs the same
- leave people unsure what the system remembers
Adding more memory makes this worse.
Correctly scoped memory fixes it.
The core idea
Intent‑Scoped Memory is an architecture, not a feature toggle.
It introduces one rule:
Past conversations should not influence the present unless a person explicitly intends continuity.
Memory is scoped by intent, not by convenience.
Memory containers
The architecture defines three distinct conversational spaces:
Channels Ephemeral attention spaces.
- No durable memory
- Context drops naturally as topics shift
- Unlimited channels without added complexity
Channels are not memory containers.
Intentional workspaces.
- Durable, isolated memory
- No leakage between threads
- Context persists across restarts
Threads are where continuity is explicitly chosen.
Shared thinking spaces.
- 1:1 DMs act as shared, durable memory
- Group DMs remain ephemeral
A core part of the architecture is study mode.
Study mode is an explicit signal that a thread is being used for real learning.
When study mode is active:
- external material is treated as primary source knowledge
- ingestion is careful and sequential
- the system slows down instead of guessing
- future answers remain grounded in what was studied
Study mode exists for the moment a person says:
“We’re actually studying this.”
Intent‑Scoped Memory defines where memory is allowed to exist. Thread checkpoints define when conversation becomes durable knowledge within those boundaries.
Without checkpoints, long threads require repeated re‑reading and reinterpretation of scrollback, increasing cost and ambiguity. Checkpoints replace implicit accumulation with explicit consolidation.
A checkpoint is a deliberate act of intent that:
-
consolidates what matters so far
-
produces a single, canonical artifact
-
becomes the stable reference for future reasoning
-
Explicit — created only by direct user intent
-
Thread‑scoped — operate only inside threads
-
Visible — surfaced back into the thread
-
Canonical — future reasoning prefers the checkpoint over raw history
-
Durable — understanding is locked in deliberately, not silently
Within thread memory, there are now two layers:
- conversational state (ongoing discussion)
- checkpointed knowledge (explicit, durable understanding)
This mirrors how people actually work: conversation continues freely, but knowledge becomes durable only when someone chooses to write it down.
Before consolidating knowledge, the system asks clarifying questions only when being wrong would matter.
It does not ask questions to reduce uncertainty, appear vigilant, or interrupt flow. Minor ambiguity is allowed to pass.
If the system ever freezes an assumption a person did not explicitly choose to lock in, it is a bug.
✅ The architectural model ✅ The mental model and invariants ✅ The rules for scoping and consolidating memory ❌ No turnkey bot ❌ No configuration files ❌ No step‑by‑step instructions
This repository defines what correctness looks like.
Intent‑Scoped Memory was designed, tested, and validated using OpenClaw as the reference implementation.
A production‑safe, step‑by‑step OpenClaw implementation guide is available separately:
👉 https://threadmind.gumroad.com/l/intent-scoped-memory
That guide covers:
- exact file changes
- session routing
- memory gating
- study mode
- thread checkpoints
- verification tests
Designed for:
- OpenClaw‑based Discord bots
- private servers
- trusted collaboration contexts
- environments where intent boundaries are socially meaningful
Not intended for:
- large public servers
- anonymous communities
- environments where consent and intent cannot be assumed
- Version: 1.1
- Primary target: OpenClaw
- Secondary applicability: Other Discord bots (with adaptation)
- Stability: Conceptually stable
- Updates: Planned
The system may notice transitions or elapsed work, but only people decide what becomes durable knowledge.
Intent‑Scoped Memory exists to ensure AI systems behave as people expect — not as databases do.