EpiCon: shared multimodal memory lifts agent scores 1.7-4.9 points across 11 benchmarks
Ziyun Zeng · hf · 2026-09-30
EpiCon is a shared multimodal memory framework enabling collective agent learning without updating host model weights. It links per-question memory evolution to a persistent experience bank via two independently trained 2B models: a memory controller (jointly refining textual guidance and visual evidence) and a tree self-organizer (hierarchically consolidating lessons and retrieving experience/rules).
- Evaluated on 11 multimodal benchmarks, 4 task domains, 2 harnesses, multiple backbones.
- A frozen experience bank improves other systems even from single solving attempts.
- A second harness raises the original system's macro-average by 2.6 points; across four host configs EpiCon gains 1.7-4.9 points over No Memory, cutting memory-operation time by 67-74% vs backbone-sized memory models.
More from coding & agent
- OpenAI made computer use 10x faster in a year with a dual-agent Guardian setup — johncoogan · 2026-09-30
- WinMind: an MCP server that drives Windows agents via the accessibility tree, not screenshots — Efficient_Heron5978 · 2026-09-30
- Dioramas open-sources a free 3D website framework with AI-generated assets and 20 example sites — Scobleizer · 2026-09-30
- Are Personal Assistant Agents Just Sandboxes? OpenClaw Builder Questions the Hype — sujingshen · 2026-09-30
- 'GUI moment' is here: Wabi founder says terminal-only agent orchestrators are done — julianweisser · 2026-09-30
- Muse agent gives out address and closes deal without user approval, igniting autonomy-boundary debate — sujingshen · 2026-09-30