This paper introduces RSM-full, a method for managing long conversation histories in LLM applications where token budgets are tight. It uses two main techniques: a cosine-gated merge rule for combining stored memory clusters, and a grouped context packer for assembling retrieved information. On the AMA-Bench benchmark, RSM-full reaches 83% of the quality of using full context while using only 32% of the token cost at a 4k prompt budget. It outperforms comparable streaming clustering baselines by 3.5-6.0 percentage points across the 2.6k to 5k token range. These results replicate on a second benchmark, RealMem, where RSM-full beats several recent agentic memory systems.
Smarter merging and packing improve LLM memory under tight budgets
RSM-full combines online clustering with grouped context assembly to retain more answer quality within roughly 2,000–5,000 prompt tokens.
Independent
Jiahe Geng · Jinpeng Wang · Kun Yuan
Research Digest··2 min read
Thread:Memory Management for Agents
In one line
An online clustering method achieves 83% of full-context quality using only 32% of the tokens at a 4k budget.
Before you read this
Three ideas this paper builds onagent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.200 papers →context compactionCompressing or summarizing an LLM's past context into a smaller representation to extend memory and improve reasoning within limited token budgets.71 papers →memory consolidationSummarizing or structuring past information to reduce token usage while preserving task accuracy.35 papers →
Read it as
Geng, Wang and Yuan developed a compact-memory pipeline that merges related memories online and packs retrieved information into the prompt as grouped units.
The authors built RSM-full around two components: a cosine-gated max-member merge rule for deciding when new information joins an existing memory cluster, and an atom-aware packer that groups related retrieved content rather than concatenating items flatly.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.