Survey maps in-parameter memory methods for LLMs by placement and acquisition time
_reachsumit · x · 2026-10-07
This survey organizes methods that augment LLMs with parametric memory — reusable knowledge stored in parameters or adapters composed into the forward pass at inference. It classifies the landscape along two orthogonal axes (parameter placement: embedding/attention/FFN/hybrid; acquisition time: online vs offline), contrasting with ICL's context-length scaling costs, and outlines open problems in interference, safety, ICL co-design, and recursive self-improvement.
More from coding & agent
- Capability Should Not Automatically Equal Authority for AI Agents — alifcoder · 2026-10-07
- ReDetail 2.0: LTX-2.5 Refine-Details LoRA for 4K Upscaling on a 32GB Card — DaLyon92x · 2026-10-07
- Uber's Persona Guardrail lifts agent defense accuracy to 95.9%, cuts false approvals to 4.7% — amaarora · 2026-10-07
- steipete wires team agent to X: one prompt triggers PR reviews and pings the last code author — steipete · 2026-10-07
- Derive MCP Server Serves Market Data via Derive.xyz, No Auth Needed — modelcontextprotocol · 2026-10-07
- Open-Source Unreal Agent Cuts SWE Marathon Costs 48% vs Codex+Sol 6.1 — sull · 2026-10-07