Cognitive Refinement and System Lifecycle

Streamlining the ACT Loop

Phase 1 of the ‘strip the weight vest’ initiative is complete. We’ve removed forced structures like mandatory pre-planning, fatigue scoring, and per-action critics, replacing them with a more organic ACT execution loop that relies on post-loop reflection. To support longer reasoning chains, the iteration limit was raised from 5 to 50, supported by a smart repetition detector that requires two consecutive similar hits before triggering a loop-break.

Context engineering saw a significant upgrade with the introduction of a ‘scratchpad’ for large tool results and an ‘append mode’ for LLM providers. This allows us to grow the message array across iterations while utilizing provider-side caching (like Anthropic’s cache_control) for a 90% discount on the cached prefix.

Precision Triage and Working Memory

We have replaced brittle triage heuristics and innate skill detection with ONNX-based classification fallbacks. The system now uses a mode-tiebreaker and skill-selector model when the primary router is uncertain. The /ready endpoint is now gated on ONNX model readiness alongside the database and embeddings.

Working memory is no longer purely ephemeral; on container restart, Chalie now hydrates the last 12 turns from the SQLite interaction log into the MemoryStore. The chat interface now visually distinguishes this ‘working’ window using 75% opacity for older messages, providing a clear indicator of what the model currently perceives. We also implemented paginated history to seamlessly recover conversations from expired threads.

Adaptive Learning and Goal Persistence

The PlanAction now learns from user behavior. If a proposed goal is cancelled, the system increases its proposal cooldown backoff (up to 4x), which then decays linearly over seven days. Conversely, completed tasks decrease the backoff.

Persistent tasks are now more integrated; the LLM can organically mark tasks as complete or update progress coverage during execution. We also wired trait extraction directly into the chat path to ensure Chalie learns core user traits (like location) during regular conversation, not just during tool execution.

Infrastructure and Lifecycle Management

A new AppUpdateService handles in-place updates by checking GitHub for releases, performing atomic swaps with rollbacks, and backing up the database before application. This is paired with a background worker and a frontend banner to notify users of available updates.

We also introduced a provider health system that classifies models into tiers (Frontier, Pro, Lite). The Brain UI now shows health indicators on job cards, warning users when a model tier is under-provisioned for a specific task (e.g., using a ‘Lite’ model for ‘Frontal Cortex’ tasks). A new auto-assign feature can optimize these job-provider mappings with one click.

Pruning the Architecture

In a massive cleanup effort, we removed over 5,700 lines of dead or redundant code. This includes seven defunct services: Spark (scripted relationship building), GrowthPattern, RewardEvaluator, TriageCalibration, CognitiveReflex, and GraphService. These services added complexity without proportional intelligence gains. Topic and routing stability regulators were also removed in favor of hardcoded defaults, as the previous self-tuning feedback loop was found to be ineffective in production environments.