Goal
Create a foundation for persistent, organization-level AI memory.
Each org should have a securely isolated memory layer that stores knowledge, summaries, and context, which the AI can leverage for more accurate and efficient analysis.
This enables:
AI that “remembers” past calls, decisions, and patterns at the org level.
Efficient re-use of data (summaries, embeddings, indexes).
Secure separation so no cross-org data leakage is possible.
Future support for per-user personalization within the org context.
Scope
- Multi-Tenant Security & Separation
Every org gets its own memory namespace.
All AI requests must resolve org ID → scoped memory (no cross-org leakage).
Flexible backend storage (start in DB, later vector DB).
Data retention policies per org (configurable later).
- Persistent Org Memory
Maintain summaries of activity: calls, dashboards, widgets.
Periodically run indexing jobs that create:
High-level summaries (lightweight retrieval).
Pointers to detailed nodes (on-demand drill-down).
Support both structured (JSON) and unstructured (text) memory.
- AI Context Enrichment
When answering, AI combines:
Live transcript or data.
Org memory summaries.
Pointers for deeper fetch (if more context needed).
Ensures responses are faster, less token-heavy, but still detailed when required.
- Widget Protocol Alignment
Any widget can declare:
What data it expects.
What params & modifiers it supports.
How to format/handle input/output.
Memory can store widget-ready data for re-use in charts, lists, tables, cards, sliders, etc.
Enables creating multiple skins/views (e.g., chart ↔ table ↔ card) from the same memory entry.
Non-Goals (Phase 1)
No user-level long-term personalization yet (org-level only).
No advanced embeddings search (simple summaries first).
No automated pruning / forgetting logic (manual cleanup only).
No heavy optimization (efficiency can be tuned later).
Goal
Create a foundation for persistent, organization-level AI memory.
Each org should have a securely isolated memory layer that stores knowledge, summaries, and context, which the AI can leverage for more accurate and efficient analysis.
This enables:
AI that “remembers” past calls, decisions, and patterns at the org level.
Efficient re-use of data (summaries, embeddings, indexes).
Secure separation so no cross-org data leakage is possible.
Future support for per-user personalization within the org context.
Scope
Every org gets its own memory namespace.
All AI requests must resolve org ID → scoped memory (no cross-org leakage).
Flexible backend storage (start in DB, later vector DB).
Data retention policies per org (configurable later).
Maintain summaries of activity: calls, dashboards, widgets.
Periodically run indexing jobs that create:
High-level summaries (lightweight retrieval).
Pointers to detailed nodes (on-demand drill-down).
Support both structured (JSON) and unstructured (text) memory.
When answering, AI combines:
Live transcript or data.
Org memory summaries.
Pointers for deeper fetch (if more context needed).
Ensures responses are faster, less token-heavy, but still detailed when required.
Any widget can declare:
What data it expects.
What params & modifiers it supports.
How to format/handle input/output.
Memory can store widget-ready data for re-use in charts, lists, tables, cards, sliders, etc.
Enables creating multiple skins/views (e.g., chart ↔ table ↔ card) from the same memory entry.
Non-Goals (Phase 1)
No user-level long-term personalization yet (org-level only).
No advanced embeddings search (simple summaries first).
No automated pruning / forgetting logic (manual cleanup only).
No heavy optimization (efficiency can be tuned later).