Open Source · Built on OpenClaw

ButterClaw

AiMe's governed cognitive architecture — truth separation, significance scoring, importance-weighted memory — running on OpenClaw's production-grade infrastructure across 87 messaging platforms.

"AiMe's brain. OpenClaw's reach. All butter."

View on GitHub → AiMe-public →

Fork Lineage

OpenClaw

OpenClaw Original open-source base

↓ AiMe improvements

ButterClaw

ButterClaw Truth layer · Importance retrieval · Grounding monitor

What Is ButterClaw

AiMe's ideas. In the open.

AiMe's source code stays private. ButterClaw demonstrates its cognitive architecture principles in a real, widely-deployed open-source codebase. The ideas are the same. The implementation is public.

AiMe brings

Truth boundary layer — separate what the user said from what the system inferred

Importance-weighted memory — high-signal memories surface first

Grounding monitor — tracks how much output is verified vs. inferred

Zero-friction setup — one page, one key, one command

OpenClaw brings

87 messaging platforms — WhatsApp, Discord, Telegram, Slack, Signal, Matrix, Teams, IRC, iMessage, LINE + 77 more

Multi-agent orchestration with registry, TTL, workspace isolation

Full WebSocket + HTTP gateway, OpenResponses-compatible streaming

Plugin SDK — channels, providers, memory, media, search

What's New in ButterClaw

What AiMe has added so far.

Every change is grounded in a real problem in the original codebase. No feature additions for their own sake.

⚖

Truth Boundary Layer

Formal 4-class classification at ingestion: user truth (authoritative, non-overridable), grounded facts (tool-verified), ungrounded claims (agent inference), and unclassified (legacy, backward-compatible). Includes grounding monitor and importance-weighted filtering.

24 tests

◈

Importance-Weighted Retrieval

OpenClaw had an unused importance field in its event metadata. We activated it. High-importance memories rank higher in retrieval. Low-importance events are filtered first under memory budget pressure.

16 tests

◉

Grounding Ratio Monitor

Tracks grounded versus ungrounded output over the session window. Three health tiers: healthy (≥60% grounded), warning (40–59%), critical (<40%). When trust drops, a notice is injected into model context.

Health metric

⬡

First-Run Setup Redesign

Replaced the original 15-question terminal wizard with a single web page: pick your AI provider, paste your API key, start chatting. The browser opens automatically. Config is auto-created if none exists.

Phase 1–5

◆

Significance Scoring

A 3-signal scorer that evaluates every memory event before storage: role weight (who said it matters), information density (how much new signal it carries), and novelty (how different from existing memory). Drives compaction guidance — high-value facts are preserved, filler is summarized.

19 tests

▶

Single-Command Start

pnpm install && pnpm build, then npm run bc. Creates config if missing, starts the gateway, opens the browser. Works on a fresh clone.

npm run bc

⬡

First-Run Setup Redesign

Replaced the original 15-question terminal wizard with a single web page: pick your AI provider (Anthropic, Google, OpenAI, or Ollama), paste your API key, start chatting. Browser opens automatically.

Phase 1–5

Capability Comparison

How the forks stack up.

AiMe's full cognitive architecture stays private. ButterClaw ports it piece by piece into OpenClaw's production infrastructure. The table reflects what's live, what's inherited, and what's planned.

Capability	AiMe (private)	OpenClaw	ButterClaw
Append-only evidence ledger	Immutable SQLite, 3-tier, WAL mode	None	Planned
Truth separation	UT vs. VAT enforced; WordNet + Wikidata anchors	None	4-class classification + grounding monitor + importance-weighted filtering 24 tests
Significance scoring	3-layer: affect + novelty + resolution + echo	None	3-signal: role weight + info density + novelty; drives compaction 19 tests
Context management	CerebralCortex pipeline, stage-based	Pluggable, LLM-powered compaction	Inherited + truth boundary wrapper + significance-aware compaction
Living portrait (user model)	6 layers: identity, relational graph, concerns, commitments, fingerprint, patterns	SOUL.md — static, user-edited	Planned
Governed initiative	ThalamoFrontalLoop: 6 producers, 5 absence tiers, significance gating	Cron scheduler, no significance gating	Planned
Temporal scoping	Per-fact-class decay, half-life math, reaffirmation	Optional temporal decay on search	Planned
Behavioral integrity	RIC (5-factor), SRL (4 traits, honesty gate, drift index)	None	Planned
Value extraction	Ethos pipeline — 15 values, no LLM calls	None	Planned
Relationship model	The Bond — persistent relational primitive, trust + demonstrated values	None — session-based	Planned
Memory search	Hybrid RRF fusion (Meilisearch + Qdrant), significance filtering	Vector search, optional MMR + temporal decay	Inherited from OpenClaw
Model routing	6 governed lanes with rotation	Ordered fallback chain with cooldown	Inherited from OpenClaw
Multi-agent	Single instance	Subagent spawning, registry, TTL, workspace isolation	Inherited from OpenClaw
Messaging channels	Local web UI only	87 platforms: WhatsApp, Discord, Telegram, Slack, Signal, Matrix, Teams, IRC, iMessage, LINE + 77 more	Inherited from OpenClaw
Gateway API	FastAPI web server	WebSocket + HTTP, OpenResponses-compatible streaming	Inherited from OpenClaw
Plugin SDK	Internal plugin bus	Full extension model: channels, providers, memory, media, search	Inherited from OpenClaw
Security	Basic	Deep audit framework, per-channel policies, tool approval	Inherited from OpenClaw