LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory