什么时候才能推出安卓预编译版

[ook826092-cloud]

希望官方能做好安卓的编译和适配

[Sanderhoff-alt]

Bringing Hindsight's memory capabilities to mobile could open up some very interesting use cases.

If uploading data to the cloud is acceptable for your architecture, leveraging Hindsight Cloud is currently the simplest and most seamless path—you avoid mobile compilation bottlenecks, storage constraints, and battery drain, while simply querying the API from the mobile client.

However, if you need on-device capabilities (for strict privacy, offline usage, or ultralow-latency local LLM setups), running the LLM and memory locally is technically possible because Hindsight natively supports any OpenAI-compatible endpoint (HINDSIGHT_API_LLM_PROVIDER=openai + HINDSIGHT_API_LLM_BASE_URL).

That said, on-device mobile setups generally fall into three completely different deployment scenarios, each with very different trade-offs:

1. Repurposed Hardware / Mini Home Server (e.g., Old Android Phone + Termux)

   • Scenario: A plugged-in, idle device with 8GB–16GB+ RAM acting as a low-power personal 24/7 server.
   • Stack: Termux running llama-server (serving quantized GGUF models like Qwen 2.5 / Gemma 2 via CPU/OpenCL) + Hindsight + embedded pg0.
   • Characteristics: Fully offline, free, low power consumption, but tailored to hobbyist/homelab setups rather than ordinary consumer apps.

2. Native Mobile App with On-Device SLM (Full Edge Autonomy)

   • Scenario: A standalone consumer Android app (e.g., a personal AI diary or assistant) where memory and model never leave the user's pocket.
   • Stack: NPU/Vulkan-accelerated mobile runtimes (MediaPipe GenAI, MLC-LLM, ExecuTorch) running 1B–3B models, with Hindsight packaged as an embedded library or background service.
   • Challenges: Mobile OS low-memory killer (LMK), thermal throttling/battery drain during frequent memory extraction, and small models (1B–3B) struggling with strict JSON schema compliance and nuanced fact extraction.

3. Hybrid Architecture (Local Data/Memory Storage + Cloud/Desktop LLM)

   • Scenario: Strict data storage privacy on mobile, without burning battery on on-device LLM generation.
   • Stack: Hindsight's database (pg0) and embeddings stay local on the phone, while heavier extraction and reflection tasks call an external LLM API (or a home desktop running Ollama/vLLM over a private tailnet/VPN).
   • Characteristics: Fast, lightweight on battery, high-quality reasoning, with raw database files remaining entirely local.

Designing how Hindsight should be packaged and adapted for Android (e.g., Termux prebuilt wheels vs. an embedded native Android runtime/library) depends heavily on which scenario is being targeted.

Could you share more about which of these scenarios you have in mind for your use case?