Hacker

Tushar Agarwal

@tushartechs

Joined Sep 2026Customer Success Engineer at GokwikIndia

Projects

  • Top 50

    SecondOrder: Theory-of-Mind Memory for AI Agents

    Graph HacksOct 2026

    Multi-agent AI systems fail at the seams. One agent promises a customer a refund, another rejects it in a channel the customer never sees, and a third picks up the conversation with no idea the customer still believes the promise. Every agent memory today (vector stores, knowledge graphs, temporal graphs) stores facts. None stores who knows which facts, and who thinks who knows them, which is exactly the context a handoff, a correction or an announcement needs. SecondOrder gives agents theory of mind as a memory layer. Every agent gets its own FalkorDB graph holding what it believes, plus nested perspectives for what it thinks other agents believe (up to 4 levels deep). Events update only the minds of the agents who witnessed them, so the graph knows who is acting on stale information, who has a wrong model of whom, and the smallest message that fixes it. Each agent can read only its own mind (FalkorDB ACLs): theory of mind, not omniscience. It's for teams building multi-agent systems (support crews, ops/incident agents, coding-agent swarms) and for organisations that need a "company brain" that knows who is out of the loop. Results: - Hi-ToM benchmark (1,200 questions, up to 4th-order belief): 100% vs the benchmark's own reference code, vs 91% for Claude Opus 5.5 alone. 87.8% vs the released labels; we found 147 labels that contradict the benchmark's own code and reported them to the authors. - Agent task: SecondOrder corrects a customer's stale belief 20/20 with a constant ~1.7k-token prompt; flat memory gets 1/20. A full-transcript handoff also gets 20/20 but needs 8.2k tokens, growing with history. - Handoffs shrink from 806 to 48 tokens. - One announcement fixes what 4 private messages can't (28 → 0 wrong beliefs about each other). - Scales to 10,200 graphs in one FalkorDB instance at 2.9 ms per event. Ships as a 5-line Python API, an MCP server for any agent, and a web app (Mind Theater, live Claude agents, plain-English event composer).