MemHarness Paper: AI Agents Should Reconstruct Memories, Not Replay Them
anselm · x · 2026-08-02
Existing memory-augmented LLM agents typically inject retrieved past experiences verbatim into the context. This often causes "negative transfer" because the static records misalign with the ever-changing current state.
Inspired by human recall, the MemHarness framework introduces a reconstructive approach. At each decision step, the agent critiques and reconstructs retrieved memories conditioned on the present context to generate grounded guidance before acting. This ability is trained end-to-end using GRPO. Experiments on ALFWorld and WebShop show that MemHarness significantly outperforms pure RL and static memory baselines, demonstrating strong out-of-distribution (OOD) robustness and fundamentally improving the agent's intrinsic reasoning capabilities.
More from coding & agent
- ERP•AI launches headless SaaS with $0 seats that bills machines: $1 buys 4,000 record writes — rohanpaul_ai · 2026-10-16
- FDE-Gym benchmark shows huge gap on real AI delivery tasks: top model scores 0.563 — DeepLethe · 2026-10-16
- Vibe-coded dashboard leaked franchise financials before anyone logged in — shauntrennery · 2026-10-16
- Fintech HR built a sourcing agent on 60,000 past applicants — 30 qualified candidates in a month — G0NZ0BOY · 2026-10-16
- Orchestrating a 100-image portrait of humanity for aliens with ChatGPT, Argon and p5.brush — dx8xb · 2026-10-16
- 3x DGX Spark cluster runs GLM-5.3-Flash locally behind a proxy, 1.75M tokens/day at 119W — Teknium · 2026-10-16