Deterministic memory layer for compliance agents: rules before LLMs, AI never changes verdicts
Friendly-Hat9209 · reddit · 2026-09-30
A cyber assessment team shares design decisions behind Hindsight, a memory layer for their compliance agent:
- Rules before LLMs: classify findings deterministically before AI enrichment.
- History changes behavior: first failure = NEWFINDING, second = RECURRINGFINDING.
- AI stays in its lane: it explains history but never changes compliance verdicts.
- Human decisions persist: exceptions and remediations become queryable memory events.
- Patterns emerge: 3+ related controls with the same finding type trigger PATTERNDETECTED.
- Memory isolation: every query is scoped by organizationid.
Their motto: "Memory adds context. Rules keep trust."
More from coding & agent
- Opus Video-Gen Workflow: Claude Code + OpenRouter to Call Every Modality With One Key — dragon_khoi · 2026-09-30
- Making games with Opus 5.5? You need to be Pinterestmaxxing first — nptacek · 2026-09-30
- Codex CLI refresh: voice-steered tasks and a new /agents view — testingcatalog · 2026-09-30
- Conductor adds Sign in with ChatGPT to bring Codex subscriptions over — charlieholtz · 2026-09-30
- OpenAI Teases Decisions API Using New gpt-6-luna Model for Fast Text and Image Decisions — stevenheidel · 2026-09-30
- Developers are already building Codex-native apps designed to run inside ChatGPT — danshipper · 2026-09-30