Memory vs computation: a Reddit design debate on keeping LLM memory out of the numbers
Strange_Wait7520 · reddit · 2026-11-08
A Reddit post explores a core architecture question for decision-support agents: should AI memory ever influence the actual computation, or only feed context to a separate verifiable system?
The author's friend's WHATIF project (small-business simulations) illustrates one answer:
- Simulation: a Bayesian model runs 5,000 simulations to produce profit/cash-flow ranges instead of a single confident prediction;
- Memory (Hindsight layer): stores reasoning behind past decisions and retrieves relevant precedents;
- Separation of duties: memory supplies context but cannot modify the simulator's numerical outputs;
- LLM constraints: the LLM only fills explanatory placeholders, and sentences containing LLM-generated digits are rejected;
- Evaluation: predictions are locked before implementation and later compared to actual outcomes; anomalous periods (festivals, stockouts) can be remembered without polluting normal training data.
The author asks other agent builders about failure modes of this separation and whether memory-computation independence is a useful design principle.
Related event: Reddit debates separating memory from computation in decision agents(2 posts)→
More from coding & agent
- Twombly launches: a shared canvas for collaborative design with AI agents — floguo · 2026-11-08
- GraphRAG vs RAPTOR vs HippoRAG: how to pick a RAG architecture for AI agents — MaryamMiradi · 2026-11-08
- Dev builds anti-hallucination research agent that verifies 108 citations word-for-word — ApprehensiveMark1308 · 2026-11-08
- Interview with Microsoft CoreAI EVP Jay Parikh on Copilot Fusion and AI code review — kimmonismus · 2026-11-08
- Agentic Creativity course recordings now publicly listed on YouTube — round · 2026-11-08
- LittlePly: a chess agent combining transformer policy, value search and bounded Stockfish help — andreisavu · 2026-11-08