Adding Persistent Decision Memory to LLM Agents via Decision Notes
adi1405 · reddit · 2026-07-06
An agent memory pattern inspired by the software engineering concept of ADRs (Architecture Decision Records) is gaining traction. Unlike vector databases or knowledge bases, this approach inserts a "decision layer" into the context pipeline: maintaining a decision-notes/ directory to record historical choices, supporting evidence, and "when to revisit" trigger conditions. This forms a judgment chain of "Information Source → Knowledge Base → Decision Record → Agent Action."
Before executing a tool, the agent checks the consistency of the decision records. If a new action conflicts with a historically human-confirmed decision, it is flagged rather than blindly executed. This pattern aims to solve the system prompt bloat and behavior drift issues in long-running agents, and is considered a lightweight alternative to the currently over-relied-upon vector DB solutions.
More from coding & agent
- WanGP adds one-click AMD GPU support for Tencent's Hy-Nomas world model, runs on 8GB VRAM — cocktailpeanut · 2026-10-27
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Founder generates 46 app promo videos in 4 days with Claude, skipping $11,500 in UGC fees — PrajwalTomar_ · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27