Older adults with memory concerns may benefit from technologies supporting physical, cognitive, and social engagement, yet these tools are often designed without sufficient input from intended users. It is also unclear how large language models (LLMs) can be incorporated in acceptable and ethical ways for this population. This study explored how older adults with memory concerns and aged-care experts perceived three digital companions, i.e., a humanoid robot (Pepper), a virtual human (Mary), and a pet robot (Paro), and identified the features and conditions that would make such technologies useful in everyday life. Ten older adults with memory concerns and ten aged-care experts each took part in an individual session involving hands-on interaction with all three technologies across physical, cognitive, social, and psychoeducational activities, followed by semi-structured interviews. Data were analysed using thematic analysis. Participants viewed the three agents as serving different roles. Pepper was associated with activity prompting and embodied guidance, Paro with comfort and emotional support, and Mary with explanation, structured conversation, and ease of use. Across technologies, engagement depended on interaction clarity, predictability, and low cognitive effort. Participants valued features that promoted independence, provided clear goals and feedback, reduced mental effort, and helped translate intentions into action. Desired capabilities included adaptive coaching, real-time feedback, and conversational continuity. Acceptance of LLM-based interaction depended on transparency, privacy protection, reliability, and clear boundaries between supportive technology and professional care. These findings highlight the importance of designing digital companions that are practical, cognitively accessible, and responsive to the needs of ageing populations.