Treasury Review Service Shows How to Use AI Memory as Context, Not Authority
joshita_09 · reddit · 2026-09-29
A developer shares a treasury review system built on a strict rule: Hindsight memory can surface relevant history, but current exposures, policy checks, and human decisions stay anchored in verifiable application records.
- Stack: FastAPI backend, Frankfurter dated FX rates, SQLite for exposures/policies/audit events, Hindsight Cloud for recall/retention, Groq for written assessments.
- Assessment flow: refuses to assess without exposures or policies; fetches market data, computes risk locally, evaluates policy rules, compares against recent historical moves, recalls memories, then sends assembled context to the reasoning model — saved with inputs and provenance.
- Memory discipline: a dedicated bank retains only sourced market assessments and human decision outcomes, tagged with metadata; recall results are app-filtered to exclude synthetic markers, with a tight token budget keeping reasoning input focused.
The takeaway: each component has a defined job — plain code calculates and enforces, the model explains, memory adds context, a person decides, SQLite preserves the app's own account.
More from coding & agent
- Designer shares AI-generated website demo, releases HTML file and prompts used — Tegadesigns · 2026-09-29
- Free Systems research: shared message boards give AI agent swarms 'shared delusions' — soumitrashukla9 · 2026-09-29
- YC's Garry Tan fixes 28 bugs in one AI-assisted 'fix wave' PR with Capy — garrytan · 2026-09-29
- Open-source workflow splits any audio perfectly for seamless MiniMax continuation — roychodraws · 2026-09-29
- Open-source Raven agent runs 42 autonomous rounds over 4 days to ship a playable FPS — lxfater · 2026-09-29
- Dev Seeks Local LLM Stack to Reverse Old Game Client, Hits Safety Guardrails — danielfrances · 2026-09-29