I'll be back. And I route tokens properly.

Before anything else happens, Clawminator classifies the request. A shared embedding (nomic-embed-text, 768 dimensions) is computed once and consumed by four parallel heads:

Downstream: Anaphora Resolver (layers A–D local, layer E optional cloud) resolves "it", "there", "that" against the dialog state. Dependency Detector (deixis + lastResult slot) detects references to previous answers. Output: a query plan with a list of classified segments.

Information retrieval is not "query one vector DB". Clawminator orchestrates four parallel sources policy-driven — how strongly each source is weighted depends on the task type.

RRF Fusion (Reciprocal Rank Fusion): Results from all sources are fused via a policy matrix. Example: for knowledge_personal, the Knowledge Graph weighs more; for knowledge_general, the Document Store does. No static merging — the weights come from the query classifier output.

The router runs through seven layers in fixed order. Each layer can match and answer directly, match and forward, or not match at all.

After the router: Confidence Gate (ADR-025). HIGH (gap > 0.02) → execute skill. MEDIUM (0.005–0.02) → Intent-LLM (qwen3.5:9b local, 1-token answer, 200–500ms) clarifies. LOW (< 0.005) → forward to Layer 3 LLM.

When a request really does need the LLM, it doesn't just go in as a prompt. Four stations before it keep the context small and relevant:

After the LLM response, a rule-based Quality Gate checks the output locally — no additional LLM call. It looks for typical failure patterns:

• too_short — response truncated or empty
• repetition — model repeats itself (infinite loop)
• placeholder — "[insert answer here]" or similar
• refusal — unwanted "I can't do that"
• lang_mix — wrong language or mixed

Mistral Improve (async Fire-and-Forget): When cloud=on and the Quality Gate flags a FAIL, the response is sent to Mistral Small in the background for an improved version. The first (local) answer goes out immediately — no extra waiting for the user. The improvement arrives as a follow-up marked with a ☁️ icon.

Not every request needs a full generation budget. The router analyzes the complexity of each message heuristically (no LLM call, <5ms) and automatically selects token budget and execution target.

Local model: qwen3.5:9b runs on Mac Mini M4 with 16 GB RAM. The trick: it's not the model that has to be small, it's the context. Through classifier + retrieval only the truly relevant context enters — 16K tokens are enough.

Four complexity tiers:

Cloud option: Currently integrated is Mistral Small with 128K context. Chosen for EU servers and GDPR compliance. More providers are planned — routing stays identical regardless of which model ends up receiving the prompt.

Hardware profiles as JSON files with schema validation. Each profile defines model selection, context budget, memory parameters, timeouts and Ollama tuning.

Currently tested:

What each profile will control: model configuration, context budget with ratios (system prompt 8–20%, memory 10–40%, history 37–45%, tools 10%), memory parameters (maxResults, minScore, fusion weights), stability timeouts and Ollama-specific tuning (keepAlive, flashAttention, kvCacheType).

OpenClaw provides the base: SQLite with FTS5 full-text search and vector embeddings. That works.

Clawminator adds: A Knowledge Graph as SQLite triple store. Two tables (entities, relations) in memory.db. Traversal via recursive CTEs — up to 3 hops. Temporal weighting and confidence-based extraction via regex + LLM.

Retrieval scoring (based on Park et al., 2023):

score = 0.35 × weight + 0.35 × confidence + 0.30 × recency

Temporal Contradiction Detection: Single-valued predicates automatically invalidate outdated facts ("lives in Vienna" replaces "lives in Berlin").

Cross-system index: KG facts are synchronized into full-text search.

Hybrid retrieval (OpenClaw): BM25 + vector cosine fusion, configurable per profile.

Memory lifecycle: Age-based expiry (default 90 days) + count-based trim. WAL mode prevents locking issues. Seed protection: seedFromWorkspace() doesn't overwrite user-set values on restart.

Anti-hallucination: Extracted entities must appear verbatim in the user message. No phantom connections.

Clawminator uses a hybrid approach: ≥90% plugin code, ≤10% core patches. OpenClaw is included as a git subtree with automated weekly sync via GitHub Actions.

This means: upstream updates from OpenClaw flow in regularly without breaking Clawminator code. The few core patches are clearly documented and isolated.

Clawminator follows an ASPICE-adapted requirements process with formal IDs, MoSCoW priorities and verification methods.

The manual test suite covers everything automation can't: Telegram messages, Spotify control, iMessage, camera triggers, Apple Notes — real macOS interactions that need a running desktop.

Clawminator provides a built-in health endpoint at /clawminator/health that shows the overall system state at a glance — directly in the browser, no extra tools needed.

What the health monitor checks: