Microsoft's PrisMem evolves agent memory per-capability, beats baselines by 10.5 points on BEAM-1M
microsoft · hf · 2026-10-07
Microsoft proposes PrisMem, a capability-driven framework for agent memory self-evolution:
- Problem: existing memory self-evolution methods use holistic evolution—deriving revision directions from mixed feedback and judging progress by overall performance—which obscures optimization directions and hides capability-specific gains offset by regressions elsewhere.
- Method: extends search guidance to individual capability dimensions, with dependency-aware capability selection (prioritizing targets with cross-capability benefits), history-guided diagnosis to refine capability specialists, and trace-guided integration that consolidates complementary gains into a unified memory program using behavioral differences on paired differential cases.
- Results: outperforms the strongest baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M respectively, effective on million-token histories.
More from coding & agent
- Claude Opus 5.5 Rebuilds a 2-Year-Old Tool Into a Polished Industrial Monitoring Panel in Minutes — karminski3 · 2026-10-07
- Your test suite costs $10 per run: the hidden CI bill of PR maxers — zeeg · 2026-10-07
- VelaTerm: Open-Source Terminal Lets Claude Code Agents Read Each Other's Chats and Orchestrate — george-lin · 2026-10-07
- Keep a 'dissent register' from day one when collaborating with AI agents — repligate · 2026-10-07
- Dev builds MCP server that rewrites Claude output in your voice, using 3-layer agent memory — nhrkhare · 2026-10-07
- What If AI Agents Remembered Like Living Systems? A Mycelial Framework for Agent Memory — repligate · 2026-10-07