01
Ingest
Authenticated events arrive from trusted sources and are appended to a durable JSONL queue.
Memory architecture
Continuity belongs to the platform, not the model.
FrankAI uses an owned memory layer so the selected AI model can receive relevant context just in time without pretending the model itself permanently remembers Ray or the business.
02
Normalize
Events keep source, kind, timestamps, privacy level, text and safe metadata in a consistent shape.
03
Index
A searchable index is rebuilt from the queue so new memory becomes available after ingestion.
04
Retrieve
Runtime search returns bounded excerpts with citations instead of dumping uncontrolled raw history.
Boundary rules
Useful memory without reckless storage.
- Secrets and credentials do not belong in memory.
- Raw bodies are excluded from public citation metadata.
- Private and restricted memories need deliberate handling before broader agent use.
- The memory layer stays model-portable.
Current live APIs: authenticated memory ingest and authenticated memory search.
Review governance