Skip to content

About

In neuroscience, an engram is the physical trace of a memory in the brain. We built Engram Edge to give Android devices a private, zero-latency multimodal memory that fits entirely inside a strict 4GB RAM budget.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

31 Commits

Folders and files

Repository files navigation

EngramEdge

In neuroscience, an engram is the physical substrate of a memory trace in neural tissue. EngramEdge brings this concept to Android edge devices—delivering a private, zero-latency multimodal memory vault and hybrid agentic assistant tailored for strict 4GB RAM budgets on tiny phones and Full Android Watches.

Instead of sending every voice memo, camera snapshot, or personal inquiry to the cloud, EngramEdge fuses EmbeddingGemma 2 (256-dimensional Matryoshka vectors) in ObjectBox 6 HNSW with Android 15 AppSearch BM25F for instant, offline hybrid recall. A zero-marginal-RAM System-1 router evaluates intent and adversarial safety in under 30 milliseconds. Orchestrated by JetBrains Koog and powered locally by compact LiteRT-LM models (including Qwen3-0.6B, FunctionGemma 270M, and Gemma 3 270M), routine memory Q&A, OS-level Jetpack AppFunctions, and native hardware skills run 100% offline. When complex real-time camera assistance or spatial Google Maps grounding is requested, the agent seamlessly escalates to Firebase AI Logic via Gemini 3.8 Flash.


Architecture & Documentation

  • System Architecture Specification: Complete engineering specification covering the 4GB memory topology, Matryoshka 256d vector embeddings, ObjectBox + AppSearch hybrid vault, Jev-style System-1 router, JetBrains Koog 1.3.0 orchestration, Jetpack AppState cross-component synchronization, and 3-tier health hardware bridge.
  • Getting Started & User Guide: Step-by-step 5-minute quickstart, micro-display ergonomics (~2.0"–2.4" screen, 80dp touch targets), built-in device skills, wireless ADB setup, and sensor tracking.

Core Pillars

Pillar Implementation Key Hardware Benefit
Form Factor & Display Micro-Deck Compose UI (sw360dp–sw426dp, AMOLED #000000) 80dp touch targets & 20–26sp text designed for small Full Android Watches and ~2" tiny phones
4GB Memory Topology AdaptiveResidencyMonitor (CO-RES default, AUTO-SWAP fallback) Stays safely under the ~1.1–1.4 GB LMKD ceiling with dynamic 4096/2048 KV-token scaling
Multimodal Vault EmbeddingGemma 2 (256d MRL) + ObjectBox 6 HNSW + Android 15 AppSearch BM25F Hybrid vector-lexical recall with Reciprocal Rank Fusion & Pre-LLM Grounding Gate
System-1 Router 256d Centroid Classifier + Temperature Softmax ($T=0.08$) Sub-30ms intent routing and confidence gating with 0 MB extra RAM
Adversarial Safety LiteRT 2.3.0 CompiledModel BERT q8 Prompt Injection Shield Dual chokepoints at ingress (OCR/Voice) and egress (run_intent/run_js)
Orchestrator JetBrains Koog 1.3.0 + KV History Compression Progressive SKILL.md tool disclosure and seamless edge-to-cloud hot-swapping
Cross-Component AppState Jetpack AppState (StateStore + PersistToDataStore + Transform) Syncs foreground UI with headless AppFunctionService & derives HUD projections off the UI thread
Sensor & Health Bridge 3-Tier BetterHealthTrackerBridge (TraX V2 ContentProvider + native Sensor 21) Zero-PPG LED battery drain with immediate sensor release
Cloud Escalation Firebase AI Logic (firebase-ai) + Firebase App Check Gemini 3.8 Flash Live 1 fps streaming & Google Maps grounding with zero Cloud Functions

About

In neuroscience, an engram is the physical trace of a memory in the brain. We built Engram Edge to give Android devices a private, zero-latency multimodal memory that fits entirely inside a strict 4GB RAM budget.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages