Agent kept misdiagnosing complaints — team shares 4 memory design decisions that fixed it
Financial-Shirt2304 · reddit · 2026-09-30
The team behind PulseMind, an agent that reads customer feedback to explain product issues, found it great at describing the present but useless at anything involving the past — it couldn't tell if an issue was new or already fixed.
The author, who built the memory side (using Hindsight), shares the design decisions that actually mattered:
- Define what counts as memory: only merged PRs and published releases are stored as "product changes"; pushes and open PRs are noise
- Separate memory types: feedback, changes, and measured outcomes are stored as distinct kinds
- Never let the LLM compute metrics: Python does the math, the model just explains it
- Query memory only when needed: it adds latency and can get rate-limited
The highlight: with memory, the agent noticed nothing in history mentioned a payments fix and flagged the complaint spike as unaddressed instead of confabulating a cause. They also added a "grounded in memory" label and a browsable memory explorer to show which memories shaped each answer.
More from coding & agent
- Adding a Mac Mini to the Cloud So Coding Agents Can Debug iOS Keyboard Bugs — lucasmeijer · 2026-09-30
- Developer flags Claude Code's memory usage as seriously bad — Sauers_ · 2026-09-30
- AI agent spent 45 minutes failing at a signup the author did in 3 minutes — paulnovosad · 2026-09-30
- Warp Terminal adds ChatGPT sign-in, letting subscribers use Codex quota in agent CLI — vikvang1 · 2026-09-30
- Dev reports sol 6.1 still finding bugs after Opus 5.5, calling it a good sign — banteg · 2026-09-30
- pi coding agent takes the stage in live demo — mitsuhiko · 2026-09-30