Developer Notes & Long-Term Persistent
Memory — Gemini Jarvis Project
1. System Hardware Constraints & Target Llama
Model Selection
The project is optimized for deployment on consumer-grade hardware with specific
VRAM limitations. The following parameters define the operational envelope for the
local LLM integration.
Component Specification / Selection
Target GPU NVIDIA GeForce RTX 3050 Laptop /
Desktop GPU
GPU VRAM Available 6 GB VRAM
Primary Model A Llama-3.2-3B-Instruct (High
speed/Low latency)
Primary Model B Llama-3.1-8B-Instruct-Q4_K_M (High
reasoning/Function calling)
Performance Baseline
● Llama-3.2-3B-Instruct: Fits entirely in ~2.2 GB VRAM, allowing ultra-fast
inference (>50 tokens/sec) and low latency.
● Llama-3.1-8B-Instruct-Q4_K_M: Uses ~4.9 GB VRAM, running fully GPU-
accelerated on 6GB VRAM while preserving high reasoning and function-
calling capability.
Scope Exclusions
Per user request, STT (voice-to-text) and TTS (Jarvis voice) modules are currently
excluded. The primary focus remains on PC automation, file management, system
control, and intelligence.
2. Core Architecture & Module Design
The system is designed as a modular framework to facilitate ease of maintenance
and clear separation of concerns.
● [Link] / jarvis_core.py: Main interaction loop, prompt caching,
Ollama API wrapper, and classification engine.
● pc_controller.py: PC automation layer (file CRUD, directory navigation,
system command execution, window/app control).
● browser_agent.py: Playwright CDP browser automation module for tab
management, scraping, and web navigation.
● memory_manager.py: Long-term memory store with deduplication and
category ranking.
● task_executor.py: Macro/sequence execution engine for multi-step PC
operations.
3. Development Cadence & Schedule Rules
To maintain the long-term memory and integrity of the codebase, the following
workflow rules are established:
Frequency
Scheduled automation occurs every 2 hours daily.
Session Workflow
1. Read Developer_Notes_and_Memory for current state and roadmap.
2. Implement new features, modules, refactorings, or optimizations in the
geminai Jarvis folder.
3. Update Developer_Notes_and_Memory with session logs and next
session roadmap.
Sunday Routine
Re-organize, audit, clean up, and restructure codebase files for high modularity and
cleanliness.
4. Change History & Log
2026-08-06 (Session 1 Initial Setup)
● Folder geminai Jarvis created.
● Developer_Notes_and_Memory initialized.
● System specs evaluated (RTX 3050 6GB VRAM).
● Llama-3.2-3B and Llama-3.1-8B-Q4 selected as primary target models.
● Architecture blueprint defined.
Section 1 Foundation Completed (Aug 6, 2026):
● llama_client.py: Built Ollama client wrapper with automatic retry
backoff, optimized for Llama-3.1-8B-Q4 (~4.9GB VRAM) and Llama-3.2-3B
(~2.2GB VRAM) on RTX 3050.
● file_manager.py: Implemented complete local file CRUD, glob searching,
and directory tree navigation.
● system_automation.py: Built process management, application launcher
registry, and system metrics collector.
● memory_store.py: Implemented persistent JSON memory engine with
Jaccard word-set deduplication.
Section 2 Expanded Foundation Completed (Aug 6, 2026):
● command_parser.py: Built regex intent classifier for parsing (:APP:),
(:URL:), (:TYPE:), (:HOTKEY:), (:PROJECT:), (:WRITE:), and
[REMEMBER] tags.
● browser_automation.py: Built Playwright CDP browser controller for
socket port checking and remote debugging navigation on port 9222.
● macro_runner.py: Implemented multi-step macro execution engine
sequencing system, file, and browser actions.
● [Link]: Built centralized configuration module for RTX 3050 6GB VRAM
memory limits, Ollama host defaults, and system directory mapping.
Next Session Roadmap (2-Hour Recurring Schedule):
● Dynamic prompt caching & instruction injection optimization in
jarvis_core.py.
● Async file watcher & directory change event listener in file_manager.py.
● Advanced Playwright web page scraper & selector interaction in
browser_automation.py.
Double Development Session Completed (Aug 6, 2026):
4. llama_client.py: Built Ollama client wrapper with automatic retry
backoff, optimized for Llama-3.1-8B-Q4 (~4.9GB VRAM) and Llama-3.2-3B
(~2.2GB VRAM) on RTX 3050.
5. file_manager.py: Implemented complete local file CRUD, glob searching,
and directory tree navigation.
6. system_automation.py: Built process management, application launcher
registry, and system metrics collector.
7. memory_store.py: Implemented persistent JSON memory engine with
Jaccard word-set deduplication.
8. command_parser.py: Built regex intent classifier for parsing (:APP:),
(:URL:), (:TYPE:), (:HOTKEY:), (:PROJECT:), (:WRITE:), and
[REMEMBER] tags.
9. browser_automation.py: Built Playwright CDP browser controller for
socket port checking and remote debugging navigation on port 9222.
10. macro_runner.py: Implemented multi-step macro execution engine
sequencing system, file, and browser actions.
11. [Link]: Built centralized configuration module for RTX 3050 6GB VRAM
memory limits, Ollama host defaults, and system directory mapping.
12. jarvis_core.py (Upgraded Main Loop): Integrated LlamaClient,
CommandParser, MacroRunner, FileManager, and MemoryStore into a
unified, high-speed execution pipeline.
13. async_file_watcher.py: Built background directory watcher for file
creation, modification, and deletion event tracking.
Roadmap for Upcoming 2-Hour Recurring Runs:
● Advanced Playwright selector interaction & web scraping macro dispatching
in browser_automation.py.
● System metrics dashboard (GPU VRAM, CPU, RAM) logging in
system_automation.py.
● Advanced macro persistence and custom workflow registry in
macro_runner.py.
Session 1: Advanced Playwright Web Scraping & Selector
Interaction
● browser_automation.py: Implemented Playwright CDP connection
handling (port 9222), automated DOM element extraction, form filling,
selector evaluation, and network idle state waiting routines. Added
automated retry wrappers for socket connections.
Session 2: System Metrics Dashboard & VRAM Monitoring
● system_automation.py: Integrated GPU VRAM (NVIDIA RTX 3050 6GB
target), CPU utilization, and RAM allocation tracking. Built non-blocking
background health diagnostics and system resource polling loops.
Session 3: Workflow Registry & Macro Persistence Engine
● macro_runner.py: Implemented multi-step macro execution engine with
JSON serialization/deserialization. Added workflow registry for custom
sequence definitions, step dependency validation, and atomic step rollback
handlers.
Session 4: Dynamic Prompt Caching & Instruction Injection
● jarvis_core.py: Refactored main message handling loop to optimize
prompt caching for Llama-3.1-8B-Q4 (~4.9GB VRAM) and Llama-3.2-3B
(~2.2GB VRAM). Implemented dynamic context injection, bounded turn-
history trimming (MAX_HISTORY = 16), and tag parsing optimizations.
Session 5: Async File Watcher & Realtime Directory Diffing
● async_file_watcher.py & file_manager.py: Implemented
asynchronous directory event monitoring for creation, modification, and
deletion tracking. Added directory tree diffing engine and glob filtering for
local workspace management.
Roadmap for Upcoming 2-Hour Recurring Runs:
● Multi-agent orchestration and task decomposition pipeline in
task_executor.py.
● Automated code self-healing and error log analyzer in [Link].
● Audio/Speech pipeline preparation for future local STT/TTS optional
integration.