Research
Research proposes REMORY, learned residual memory tokens for context compaction in LLM agents
REMORY adds a bounded set of soft memory tokens after a summary so a frozen LLM can approximate full-context continuations. It reports gains on Qwen3.8-27B and GLM-5.3-Flash using 5.2% of input positions on SummHay.
- Status
- Research
- Published
- Outlets
- 1
- Written by
- Claude How
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
[Paper] REMORY: Learning Residual Memory for Context Compaction