Jun 2026· arXiv.org· Vol abs/2606.29914· 2 citations· 20 references
Computer Science
TL;DR
MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families), recommends memory evaluations fix embedding models across comparisons, stratify by model family, and report write-path cost before attributing gains to architecture.
Abstract
Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured. We present MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families). Four findings emerge: (1) verbatim RAG matches full-context GPT-4o-mini (47.2% vs. 49.8%, p = 0.34), but the ranking reverses across models: Gemini gains +14pp from full context, while Sonnet gains +31pp from RAG, partly because it refuses 63% of full-context queries; (2) swapping only the embedding model in an identical pipeline shifts accuracy by +6.2pp at n = 500 (p = 0.004), and Mem0 beats MiniLM-RAG by +11pp but loses to cloud-RAG by 1.2pp, so one variable flips the conclusion; (3) agent self-memory (42%) underperforms basic retrieval (47%); (4) on 2 of 6 question types (n = 88), Mem0 matches cloud RAG (72.7% vs. 73.9%, p = 1.0) at 50x the cost, suggesting narrow rather than general gains. We recommend memory evaluations fix embedding models across comparisons, stratify by model family, and report write-path cost before attributing gains to architecture.
Most agent-memory benchmarks test post-hoc recall, whereas MemoryArena evaluates whether memory supports interdependent, multi-session task completion. We compare MemoryLake, a structured multi-track memory backend, with Mem0, text-embedding-3-small vector RAG, and a long-context control across all five MemoryArena domains. The systems share the same agent framework, requested gpt-5-mini model alias, task samples, and scoring code; the memory integration is the intentionally changed component. Because each backend bundles write, retrieval, consolidation, budgeting, and prompt-assembly choices, the study is a matched system-level comparison, not a representation-only ablation or a cost-matched experiment. On the shared evaluation sets, MemoryLake has the highest observed success rate (SR) in mathematics (9/40), physics (12/20), and progressive retrieval (4/20). Every system has zero SR in travel planning, and web shopping yields a single bundle-level success (long context, 1/150); MemoryLake ranks third on both the travel soft process score and shopping step match. Following MemoryArena's suite-level convention, a post-hoc equal-weight average over the five SRs is 20.5% for MemoryLake versus 13.6% for the best comparator. These are point estimates: sample sizes are modest, confidence intervals overlap, and we do not report paired significance tests. A separate MemoryLake-only run over all 221 progressive queries yields a failure-counted SR of 26.7% (59/221) and is not a baseline comparison. The results support a workload-dependent view of memory backends and an observed lead among the four evaluated systems on the shared sets; they do not establish benchmark-wide state of the art or a causal advantage of representation structure.
Chao-Shun Zhan, Qiang Zhou, Guannan Li et al.· 0 citations
Agent-memory workloads mix direct factual lookup, relation-chain and current-state reasoning, and broad synthesis over long histories. We describe Supra Cognitive Modes (SCM), an architecture that maps explicit or automatically selected per-query modes to retrieval and synthesis payloads over one shared ingest substrate. A frozen semantic classifier and runtime gates dispatch queries among fused lexical and dense lookup, graph or iterative multi-hop handling, and stratified long-form synthesis. The substrate combines multi-granularity embeddings, extracted triples, fact-version metadata, and optional asynchronous enrichments. We characterize the deployed configuration on three benchmarks: Long-term Conversational Memory (LoCoMo; n = 1,986), MemoryAgentBench (MAB; n = 3,671), and LongMemEval (n = 500). The reference run records 84.87% on LoCoMo factoid categories and 68.61% on adversarial abstention, 61.49% on MAB across two repetitions, and 86.00% on LongMemEval. A repository-backed reproduction produces similar aggregate scores and supports task- and mode-conditioned failure analysis. Raw baseline outputs, aligned end-to-end timing for LoCoMo and LongMemEval, and complete token ledgers are unavailable; stored rows also omit some final runtime decisions. The results characterize one implemented routed configuration and its diagnostic failure patterns, while source inspection verifies the per-query control interface and shared-substrate design. Causal routing effects, efficiency gains, and statistical significance remain outside the available evidence.
Memory Manager models are pivotal in agent systems. Existing reinforcement-learning methods commonly use LLM-judged synthetic question-answer (QA) pairs: this provides useful downstream task grounding, but values memory through a sampled query distribution and a fixed reader. We propose CMI-Mem, a lightweight RL memory manager with a hybrid reward. Its extrinsic QA term measures end-task correctness, while its intrinsic Conditional Mutual Information (CMI) term evaluates the information contributed by new conversational inputs relative to the current memory state without conditioning on a sampled QA query. The two signals are complementary: QA anchors task utility, whereas CMI provides per-operation supervision for relevant, non-redundant memory construction. Experiments demonstrate improved transfer across memory-use scenarios, together with more efficient training and inference from the per-operation CMI signal. Our codes are available at: https://github.com/Wyb0627/CMIMem , and the CMI-Mem-4B model checkpoint is available at: https://www.modelscope.cn/models/wyb0627/CMIMem-4B
Yubo Wang, Qiuyu Zhao, Zenghui Sun et al.· 0 citations
MemDefrag, a training-free and model-agnostic framework that uses a middle-layer tracing signal to conduct memory defragmentation (rank, reorder, and filter memories), and applies an informativeness-guided proportional forgetting mechanism once capacity is exceeded, is proposed.
Memoir combines per-sample fast memory, shared slow parameters, variable-depth latent recurrence, and a future-latent energy objective. We test its riskiest coupling: each pondering iteration may rewrite the fast tier that the same iteration reads. On procedural associative recall with key interference, we compare a coupled arm against an otherwise identical read-only pondering arm. Both arms contain 81,738 parameters, including 76,362 trainable parameters, and use matched declared forward multiply-accumulate counts, data, optimizer, schedule, and seeds. After 240 training steps across 12 seeds, coupled recall is 0.5203 with a 95 percent interval of [0.4522, 0.5883], while read-only recall is 0.6557 with [0.5953, 0.7160]. The arms are paired per seed, and the read-only lead of 0.1354 gives a paired t of 3.23 on 11 degrees of freedom with a 95 percent interval of [0.0431, 0.2277] on the difference, winning on 10 of 12 seeds. After 960 steps across 8 seeds, both arms reach 1.0000, so the measured effect is a learning-speed penalty at a fixed budget, not a demonstrated capability penalty. That longer control is ceiling limited, leaving convergence on a non-saturating task unmeasured. A predicted failure in which memory rewriting corrupts the energy signal did not occur: the energy margin grew and held. Kernel restructuring also reduced delta-rule forward time from 0.907 ms to 0.351 ms on the stated device. Code and evidence are available at https://github.com/RightNow-AI/Memoir
Long-term memory is becoming a core component of LLM agents, but most memory benchmarks evaluate conversations or compact summaries, while research agents need to restore evidence from full scientific papers. We introduce two full-text scientific-memory benchmarks, Public AI Memory (PAIM; 81 papers, 66 questions) and Public Transformers (PTr; 252 papers, 98 questions). We evaluate eight memory/retrieval systems, including our own proposed Theoria, plus a no-retrieval baseline. Our results show that memory leaderboards are not interpretable without the full protocol: ingestion granularity, raw-text preservation, retrieval budget, retrieval modality, rubric audit, and judge choice all affect the outcome. For example, on PAIM Graphiti wins convincingly but uses 2.6M characters of retrieved context per query, and after controlling for retrieval budget the lead disappears. On PTr, for the systems where BM25 retrieval can be added cleanly, the sparse-dense hybrid is the single most significant intervention: hybrid variants of Simple RAG, Mem0, and Theoria tie for the lead within 0.03 points. Multi-judge and human side-by-side calibration show that LLM-as-a-judge rankings are consistent across frontier judges and agree with human evaluation, with an effective resolution of roughly one point on a ten-point scale. We argue that scientific memory should be evaluated as budgeted, modality-aware context restoration rather than as an unconstrained architecture leaderboard, and we release the datasets, harness, raw outputs, judgments, and scripts to reproduce our results and serve as tools for such evaluation. Our code is available at http://gitlab.com/quantellence/research/scientific-recall-bench , and the datasets are available at http://huggingface.co/datasets/quantellence/srb-data .
Maksim Sheverev, David B. Finkelstein, Sergey I. Nikolenko· 0 citations