ARCHITECTURE

TR-2025-08

Adaptive Cognition Layer Pipeline

The ACL orchestrates ten sequential components that transform user input into contextualized responses through intent classification, memory retrieval, reasoning routing, safety validation, and model selection.

10

Pipeline Stages

96.7%

Local Routing Success

94%

Hallucination Detection

0.73

Confidence Correlation

COGNITIVE PIPELINE

01

Intent Classifier

Categorizes user input by intent type: question, task, conversation, action request. Maps intent clusters for pattern analysis.

OUTPUTS

  • Intent type
  • Intent clusters
  • Routing signals
02

Context Engine

Builds app context from workspace state. Integrates active projects, tasks, notes via Contextual Priority System (CPS).

OUTPUTS

  • CPS weights: recency(0.3), frequency(0.25), connections(0.25)
  • Workspace objects
03

Workspace Memory

Retrieves relevant workspace items via RecallIndexEntry. Surfaces items with emotional snapshots (valence, tone, intensity).

OUTPUTS

  • Recall entries
  • Emotional snapshots
  • Importance scores
04

Long-term Memory

Accesses conversation history and semantic memory. DBSCAN clustering for theme discovery across sessions.

OUTPUTS

  • Memory graph clusters
  • Cross-conversation patterns
  • Narrative coherence
05

Reasoning Layer

Routes complex queries to Python brain modules. Triggers on: query length >80 chars, analytical patterns, explicit reasoning signals.

OUTPUTS

  • Reasoning chains
  • Module routing
  • Depth metrics
06

Safety Layer

Validates responses for hallucinations and factual accuracy. Performs recall-based fact verification against workspace knowledge.

OUTPUTS

  • Hallucination flags
  • Fact verification
  • Confidence signals
07

Personality + Style

Adapts response tone via StyleAnalyzer and ARTE. Matches user typing patterns, energy levels, and formality.

OUTPUTS

  • Tone calibration
  • Style signals
  • Personality adaptation
08

Response Construction

Builds full prompt with all context layers. Integrates workspace context, memory, conversation history, and cognitive metrics.

OUTPUTS

  • Assembled prompt
  • Context integration
  • System state
09

Confidence Scoring

Calculates self-aware confidence metrics (0.0-1.0). Based on recall quality, context freshness, and intent signals.

OUTPUTS

  • Confidence score
  • Certainty indicators
  • Transparency metrics
10

HybridBridge Routing

Selects optimal model (local/cloud/hybrid). Implements stickiness, casual detection, and thinking mode activation.

OUTPUTS

  • Model selection
  • Routing decision
  • Inference path

MEMORY SYSTEMS

Workspace Memory

RecallIndexEntry tracks all workspace objects with emotional snapshots. 78% hit rate for relevant workspace items.

Emotional dimensions: valence, tone, intensity
Importance scoring: access patterns + manual boost

Long-term Memory

Memory Graph with DBSCAN clustering. 82% accuracy for theme discovery across conversations.

Compression threshold: 40 messages
Retention: last 12 messages + compressed history

MODEL ROUTING

Local-First

Privacy-preserving operation via Ollama. 96.7% local routing success rate.

Default: qwen3:1.7b

Cloud Fallback

Automatic failover to cloud on local errors. 94.2% cloud routing success.

Cloud: gpt-oss:20b

Model Stickiness

Maintains same model for 3 consecutive turns to ensure consistency.

Health TTL: 5 minutes