Memoria V4.5: 82.6% Recall on 4GB RAM CPU-Only Machine
kitkatz69 · reddit · 2026-08-27
Memoria V4.5 achieves 82.6% Recall@1 on LongMemEval-S while running locally on a 4GB RAM CPU-only machine without a cross-encoder. The system features a modular architecture with replaceable retrieval workers, a declarative scheduler, and extensive plugin support (39 hooks). The author is seeking feedback on the retrieval/ranking boundary and scheduler abstraction.
More from coding & agent
- Agent Plugins released to foster open Agentic AI ecosystem — romainhuet · 2026-08-27
- Code review in 2026: AI transforms the workflow — tekbog · 2026-08-27
- Developer ships open-source MCP server to run Labnana image generation inside Claude Code and Cursor — e10t · 2026-08-27
- ComfyUI-Labnana nodes bring image generation, editing, 4K jobs and credit checks into ComfyUI workflows — e10t · 2026-08-27
- Terminal agent frontier shifts with harder evaluation and verifier fixes — abeirami · 2026-08-27
- Anthropic's AI-native SDLC playbook replaces line-by-line review with gate approvals — Forward_Mind6886 · 2026-08-27