Developer Builds Bi-Temporal Memory Engine to Fix AI Agent Context Degradation

A developer identified a core flaw in current AI agent design: large context windows do not prevent memory degradation, as standard vector databases cause agents to hallucinate when faced with contradictory information. To address this, they built PaperPlanes, a stateful research assistant powered by a bi-temporal memory layer using CockroachDB, AWS Bedrock, LangGraph, and FastAPI. The system never deletes knowledge; instead, conflicting facts are flagged as disputed, and a time-slider UI lets users query exactly what the AI believed at any past point. Concurrent write conflicts were handled via CockroachDB's serializable isolation, achieving zero lost writes across 25 simultaneous agent writers in stress testing. The project also incorporates an Ebbinghaus decay algorithm so that infrequently accessed memories fade while recalled facts are reinforced, and it has been submitted to the CockroachDB × AWS Hackathon.