An agent working for weeks should not forget what it learned yesterday. Long Horizon Memory gives agents persistent, connected memory so they can recognise patterns, recover earlier evidence, and keep pursuing a goal as information grows. New observations enter a short-term graph, then merge into durable memory with timestamps, source references, and contradictions intact. Semantic search finds relevant memories. Graph traversal connects the evidence. Louvain clustering groups related concepts into higher-order summaries, keeping the original sources accessible. The agent retrieves what matters within a fixed context budget. Our demonstration tackles recurring NYC 311 complaints. Across chronological batches, the agent connects reports, identifies persistent issues, and recommends upstream investigations backed by cited evidence. Complaint recurrence supplies a measurable feedback signal for evaluating recommendations and developing future learning policies. Custom Python loops orchestrate the agent. The architecture combines MongoDB Atlas, Voyage embeddings, Atlas Vector Search, NetworkX, OpenRouter, and a Next.js dashboard that exposes memory activity and decisions. We compare the harness against a sliding-window baseline to test whether older evidence remains useful after it leaves context. The ambition is coherent memory across billions of tokens. This prototype makes that ambition testable, with scalable retrieval and automated feedback learning as the next milestones.
Team
+1Placement
Participant
