Erina Karati introduces Project Paradox, a modular framework developed at Supercell's AI Innovation Lab for game agents with individual memoriesAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts., emotions, beliefs and plans. Short interactions worked well, but longer runs exposed failuresA long-horizon agent pursues an objective across many actions or extended periods, requiring reliable task state, feedback and stopping conditions.: rumors became facts, sources disappeared and remembered information did not reliably influence actions.
Erina Karati proposes an external experiment loop that runs controlled scenarios, records traces and changes only a small policy surface. A balanced scorecard measuresAgent evaluation tests whether an AI agent completes tasks correctly, consistently, and within its required boundaries. information spread, source retentionData provenance records where information came from, how it changed, and which people, systems, or processes handled it., uncertainty, action consistency and privacy. Optimizing just one score can reward oversharing or stale memories, so a change is retained only when the relevant guardrails also hold.
Erina Karati emphasizes freezing the harness, scenarios and metrics while testing memory, retrievalRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task. and communication policies. A reported improvement in one rumor scenario does not establish general improvement across the system. She presents the approach as an engineering pattern that may also help support, research and workflow agents maintain reliable state over time.
Watch on YouTube
