AuthorsJiahe Geng, Jinpeng Wang, Kun Yuan
TL;DR
RSM-full uses a cosine-gated max-member merge rule and atom-aware grouped packing to reach 0.311 (83% of Full-Context) at 32% token cost on AMA-Bench.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Full-context prompting preserves quality but its token cost grows linearly with interaction length, making it impractical beyond compact budgets ≤5k tokens.
Under tight prompt budgets, long-horizon LLM agents must drop past context, which hurts answer quality, causal reasoning, and state updates on benchmarks like AMA-Bench.
HOW IT WORKS
RSM-full centers on Cosine-gated max-member merge for streaming writes and Atom-aware grouped context assembly for compact prompt construction.
You can think of RSM-full like a smart card catalog: it clusters related experiences into atoms, then pulls grouped folders instead of scattered pages when answering.
This max-member gate and grouped packer let RSM-full keep high-quality, structured memories within 2k–5k tokens, something a plain context window or flat retrieval cannot do.
DIAGRAM
This diagram shows how RSM-full writes incoming chunks into atoms with cosine-gated max-member merge and later reads them with atom-aware grouped assembly.
DIAGRAM
This diagram shows how AMA-Bench evaluates RSM-full against Online K-Means, DP-means, and flat concatenation in a 2×3 factorial ablation.
PROCESS
Cosine-gated max-member merge
RSM-full encodes each trajectory chunk h, finds m* = arg max cos(h, μm), and uses the cosine-gated max-member merge rule with threshold τ≈Q0.70.
Atom-aware grouped context assembly
At query time, RSM-full retrieves top K atoms, ranks members within each atom, and packs chunks grouped by atom with headers and temporal ordering.
|v⊤₁,m q| retrieval rule
RSM-full scores each atom m using |v⊤₁,m q|, where v₁,m is the top singular vector of the atom’s member matrix, equivalent to normalized-centroid cosine on BGE.
AMA-Bench compact-memory operating point
RSM-full runs GPT-4o-mini agents under a 4k-token budget, then llmmelon GPT-5.4 judges all 2,496 QAs to measure average correctness and quality–token trade-offs.
KEY CONTRIBUTIONS
Strong compact-memory operating point
RSM-full reaches 0.311 average correctness on AMA-Bench, delivering 83% of Full-Context quality at 32% of the 12,519-token cost at a 4k budget.
Mechanism account via merging and packing
RSM-full’s cosine-gated max-member merge rule adds +5.74 pp over Online K-Means and +5.45 pp over DP-means, while atom-aware packing adds +5.02 pp over flat concatenation.
Replication on RealMem persona memory
RSM-full achieves 0.4684 accuracy on RealMem, improving Budget-RAG by +0.69 pp and Streaming-Proto by +2.97 pp under a 4k-token budget.
RESULTS
Avg (per ep) Recall
0.311
+1.9 pp over Budget-RAG at ~4k tokens
TotalTok
4,001 tokens
32% of Full-Context’s 12,519 tokens on AMA-Bench
Avg correctness RealMem
0.4684
+0.69 pp over Budget-RAG and +2.97 pp over Streaming-Proto
Clustering ablation gain
5.74 pp
RSM-full vs Online K-Means with |v1 q| retrieval on AMA-Bench
On AMA-Bench, which tests long-horizon tool-use and reasoning, RSM-full achieves 0.311 average correctness at 4,001 tokens versus Full-Context’s 0.373 at 12,519 tokens. On RealMem, which probes persona memory over 135–276 sessions, RSM-full’s 0.4684 accuracy confirms that compact clustered memory can preserve quality under 4k-token budgets.
BENCHMARK
On AMA-Bench, which tests long-horizon tool-use and reasoning, RSM-full achieves 0.311 average correctness at 4,001 tokens versus Full-Context’s 0.373 at 12,519 tokens. On RealMem, which probes persona memory over 135–276 sessions, RSM-full’s 0.4684 accuracy confirms that compact clustered memory can preserve quality under 4k-token budgets.
BENCHMARK
Average correctness on AMA-Bench at the compact-memory 4k-token operating point.
BENCHMARK
Three-seed mean accuracy on RealMem under a 4k-token budget.
KEY INSIGHT
On BGE-normalized AMA-Bench, swapping |v⊤₁,m q| retrieval for centroid cosine on RSM-full clusters changes average correctness by only +0.33 pp (p=0.40).
This is surprising because a more sophisticated singular-vector retrieval rule seems strictly better, yet RSM-full’s gains come almost entirely from merging and packing rather than retrieval scoring.
WHY IT MATTERS
RSM-full shows that carefully designed streaming clustering and atom-aware packing can define strong quality–token Pareto points for long-horizon agents.
Builders can now deploy LLM agents that retain rich, structured memories over many sessions while staying within 2k–5k prompt tokens, avoiding the costs of full-context prompting.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…