v0.4.4

v0.4.4

Memory precision, code-aware retrieval, policy hardening, and CLI resilience.

Highlights

• Exact memory provenance — memories now link to the specific source chunks they came from instead of entire ingestion batches.
• Blackboard-first memory pipeline — candidates are deduplicated and reviewed before entering long-term memory, reducing memory pollution.
• Code-aware retrieval — parser-backed code chunking, symbol graphs, import resolution, and memory_find_related for cross-file code exploration.
• Memory that learns — new lessons, skills, and reflection systems allow BrainRouter to reinforce successful patterns and extract reusable workflows.
• Graph Intelligence — PageRank, bridge detection, shortest-path analysis, and dashboard intelligence views.
• Unified execution policy — all tools, delegation, workers, and MCP calls now follow the same permission and audit system.
• Brain tree improvements — topic-aware memory trees and global rollups for higher-level knowledge organization.

CLI Improvements

• /reload-skills without restarting
• Optional edit → verify loop after file edits
• LSP navigation (definition, references, hover, symbols)
• Crash recovery and offline prompt queue
• Automatic replay of queued prompts on reconnect
• Update notifications for newer releases

Reliability

• Fixed child-agent failures causing CLI hangs
• Improved background worker isolation
• Added end-to-end redaction coverage
• Hardened tracing, path validation, and external access controls

Status

• 476 brain unit tests
• 106 integration tests
• 944 CLI tests

Full Changelog: v0.4.3...v0.4.4

Full Changelog: v0.4.3...v0.4.4