I’m building Evopien, a conversational companion prototype on Jetson AGX Thor, and wanted to share the memory architecture behind its latest demo.
Local Qwen handles conversation. Evopien Core owns memory and permission decisions. PostgreSQL is the canonical store, while Hindsight supports retrieval and does not replace Core as the authority over durable records.
The model receives filtered context rather than direct database access or permission to commit permanent memories. The memory work covers retention, later retrieval, corrections and forgetting.
One difficult part is consistency between the record and the answer. Updating a preference in the database is only useful if later retrieval and conversational context stop reinforcing the old version. A deleted or revoked memory needs the same attention.
The prototype is being developed for English and Spanish. Correction edge cases and the latency added by memory operations are still active areas of work.
Demo:
EVOPIEN | An AI Companion Robot with Memory You Control — Full Governed Memory Demo
I’d welcome technical feedback on evaluating this boundary. What tests would you use to separate a correct memory operation from a model that merely says it has remembered or forgotten something?