Reasoner-1 March 2026 Update: 360 Analytics, Hybrid Charts, and Memory Roadmap
This month we've shipped updates that make Reasoner-1 more effective for 360 analytics—period comparisons, diagnostic questions, and workflow visualizations. We're also sharing our roadmap for a memory system that will turn Reasoner into an AI analyst that learns over time.
360-Style Analytics Suggestions
When you load a dataset with date and metric columns, Reasoner now suggests high-value prompts directly from your schema:
- Trend of [metric] over last 12 months — Time series by month
- Compare [metric] vs previous period — Period-over-period comparison
- Why did [metric] drop last month? — Diagnostic with top contributors
- Top contributors to [metric] change — Driver analysis by dimension
These use real column names (revenue, region, order_date) instead of placeholders. The suggestions appear both on initial load and as follow-ups in the thread, so you can go deeper without rephrasing.
Semantic Layer in Action
The semantic layer is now fully wired end-to-end. When you configure column roles, synonyms, and default aggregations (via "Improve chat" or "Configure semantic layer"), that metadata flows into:
- Insight API — Schema hints include role, synonyms, and default_agg for better NL→SQL matching
- Reasoner — Entity resolution uses synonyms and display names for metric/dimension resolution
If you say "sales" and your column is net_revenue, adding sales as a synonym helps. Auto-inference runs when data loads; manual overrides refine it.
Memory Roadmap: Why It Matters
Today's Reasoner is stateless: each question is independent. To support follow-ups like "Which reps caused the decline?" (when you already asked "Why did revenue drop?"), we need memory.
We've designed a simple five-layer memory system:
Working Memory
In-process during a workflow: question, resolved entities, step results, findings. Already exists as EvidenceStore; we'll formalize it as a unified struct.
Session Memory
Recent questions and summaries for the conversation. Enables "that decline" to refer to the previous answer. Lightweight—Redis or in-memory.
Semantic Business Memory
Dataset-level: metric catalog, dimension catalog, time column. Persists when the dataset loads. Already partially covered by semantic_layer; we'll add persistence.
Pattern Memory
Episodic learning: "Chennai has been a recurring revenue driver in previous declines." Save patterns after workflows when impact_score > threshold.
Procedural Memory
Tool registry: which tools solve which intents. Already implemented in Reasoner—we can expose it to the planner for dynamic tool selection.
Each layer has a clear job. We won't mix them. The goal: an AI data analyst that learns over time, not just a chat assistant.
What's Next
We're implementing Working Memory formalization and Session Memory (client-side first) next. Pattern Memory will follow once we have a backend store. Semantic persistence will round out the dataset knowledge layer.
Try the new 360 suggestions and workflow charts in Chat with Fluxor. If you have feedback or questions, we'd love to hear from you.