New MCP tool and portal UI for executing natural language instructions
against the memory graph via Ollama (qwen3-coder:30b). Single LLM call
generates a JSON action plan which is executed sequentially.
Supports 8 action types: add_node, update_node, remove_node, add_edge,
remove_edge, bulk_tag, reorganize, query. Actions can reference previous
results via $result[N].field interpolation. Uses /api/chat with few-shot
assistant example, format:json, and temperature:0 for reliable output.
- Query bar with organized/grouped search results in portal
- Maintenance panel UI for triggering and viewing maintenance status
- Heartbeat service with periodic maintenance and dirty-tracking
- Query organizer for grouping search results by tag/kind/parent
- Slide-up animation for query panel
Add lastAccessedAt timestamp to nodes with schema migration and
backfill. Touch timestamp on read, apply exponential freshness decay
(~69-day half-life) to search scoring alongside BM25 and vector
weights. Add auto-decay that marks untouched nodes as stale after a
configurable threshold, with CLI command and server-side daily interval.
Include comprehensive USAGE.md documenting all CLI commands and REST API.
Replace fixed sidebar with floating action buttons (menu + add node) in
bottom-left corner. Sidebar and node panel now slide in as overlay drawers
with backdrop dismiss, capped width for small screens, and slide animations.
Express API server wrapping existing store/graph core with REST endpoints
for nodes, edges, graph, and search. React + Vite portal with React Flow
for interactive graph visualization, Tailwind CSS styling, and full CRUD UI
(sidebar, node panel, add/link modals, search bar, toast notifications).
