Major Memory & Tooling Overhaul

A complete memory recall overhaul introduced a new tool output format of [id:{id},relevance:{level}] {value} and added a reflect action featuring deep episode search and 3-level graph expansion.

The data graph architecture was unified, replacing the legacy knowledge table with a service that uses ACT-R activation and a split strength model for storage and retrieval.

Memory skill logic was rewritten to interact with the new DataGraphService, centralizing contradiction handling and using composite scoring for recall.

The system underwent extensive cleanup, ripping out several legacy components including the webhook infrastructure, digest worker pipeline, and procedural learning chains.

Tool dispatch was unified, routing all innate and external tools through a single handler path on ActDispatcherService, and search now returns structured JSON results.

The chat history mechanism was simplified by rewriting it to a flat message list, removing previous reliance on channel-based exchanges.

Significant code removal occurred by deleting services related to working memory, mode routing, identity, and various legacy prompt/skill definitions.

Database operations were streamlined, notably simplifying the privacy delete-all function to truncate only the three core tables: tool_calls, episodes, and transcript.

• Unified data_graph store replacing legacy knowledge table

• Added reflect action with recursive 3-level graph expansion

• Unified tool dispatch via ActDispatcherService

• Rewrote chat history to a flat message list CRUD

• Ripped out webhook infrastructure and digest_worker pipeline

• Removed WorkingMemoryService and ModeRouterService