How a 32GB VRAM local LLM user treats context as a scratchpad, not working memory
Training-Ruin-5287 · reddit · 2026-09-25
A Redditor shares a context-management approach for local LLM setups: treat context as a temp scratchpad for each step rather than working memory. Their rig pairs a 32GB VRAM main PC with an older 8GB machine running a 9B Qwen model in the background for compaction and long-term-memory sorting. The main model's working state lives outside the context window in continuously edited docs; context holds only the current task plus retrieval from those docs, with git providing recall and history. They're polling other low-end local hosts on how they handle context without bloated supporting systems.
More from coding & agent
- OpenAI Codex community holds first Mendoza event with GPT-6 Astra build sessions — paw_lean · 2026-09-25
- Where an Agent Loop Actually Runs, How It Fails, and Making Sessions Durable — tempNull · 2026-09-25
- Grok Bot vs Meta Muse: Crew of 50 Bots vs One Personal Agent, Which AI Employee Wins? — MicahBerkley · 2026-09-25
- Codex CLI v0.157.0 adds GPT-6 Sol and Luna with Bedrock support — github-actions[bot] · 2026-09-25
- Claude 5.5 Opus Builds a 2D/3D Video Entirely in Claude Code, No Assets or MCP — Silver-Chipmunk7744 · 2026-09-25
- Open-Source Slack-Style Workspace Where Humans and AI Agents Build Together Goes Viral — Saboo_Shubham_ · 2026-09-25