AuthorsYasmine Omri, Ziyu Gan, Zachary Broveak et al.
TL;DR
Agent Memory uses a four-axis taxonomy and phase-aware profiling harness to show that construction energy for stateful agents can exceed query energy by over 47× per correct answer.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents on LongMemEval_S_* face interaction histories of approximately 360K tokens per sample, making full-history prompting both slow and energy-intensive.
Long-context baselines reach around 38 seconds per query, while Agent Memory shows that some external memory systems pay over 15,429 kJ and 47× more energy per correct answer than BM25.
HOW IT WORKS
Agent Memory introduces a system-oriented taxonomy over construction, storage, retrieval, and mutability, and a phase-aware profiling harness that tags memory construction, retrieval, and generation costs.
Think of Agent Memory as a profiler for the agent’s "RAM vs disk" split: short-term working memory vs persisted agent memory across paradigms like long-context, flat RAG, structure-augmented RAG, and agentic control flow.
This KEY_MECHANISM lets Agent Memory expose how designs like Mem0, GraphRAG, and Letta shift work from prompts into external memory, revealing lifecycle energy, latency, and footprint tradeoffs that a plain context window cannot show.
DIAGRAM
This diagram shows how Agent Memory structures the end-to-end execution pipeline from interaction ingestion to memory maintenance for stateful agents.
DIAGRAM
This diagram shows how Agent Memory’s taxonomy organizes ten representative memory systems into four paradigms by construction, storage, retrieval, and mutability.
PROCESS
Ingestion
Agent Memory ingests the interaction stream, defining memory construction units such as chunks, conversation turns, or complete sessions to control write granularity.
Memory construction
Agent Memory applies deterministic, LLM-mediated, or agentic construction to transform raw history into semantic, episodic, or procedural memory entries stored in the agent database.
Retrieval and prompt assembly
Agent Memory runs retrieval pipelines like lexical top-k, dense search, graph expand, or tool-mediated routing, then assembles relevant entries into working memory for the LLM.
Maintenance
Agent Memory performs maintenance via deduplication, consolidation, forgetting, compression, or re-embedding to manage long-term memory footprint and freshness over time.
KEY CONTRIBUTIONS
System-oriented taxonomy of agent memory paradigms
Agent Memory classifies ten systems into long-context memory, flat RAG memory, structure-augmented RAG memory, and agentic control flow along construction, storage, retrieval, and mutability axes.
Phase-aware profiling harness for agent workloads
Agent Memory introduces a profiling harness that tags every LLM and embedding call with construction, retrieval, or generation phase, recording tokens, latency, GPU utilization, and energy.
Comprehensive systems characterization and recommendations
Agent Memory characterizes construction cost shape, capability thresholds, amortization structure, freshness scheduling, footprint growth, and retrieval tail behavior, and derives ten deployment recommendations.
RESULTS
LLM-Judge Accuracy on LongMemEval_S_* (BM25)
47.0 %
+7.2 pp over Mem0 on LongMemEval_S_*
Mean QA wallclock per query (Mem0, local Qwen3-32B)
0.10 s
Mem0 vs long_context at ~38 s per query
Total energy for BM25 (Construct + 300 QA)
582 kJ
BM25 vs Letta at 15,429 kJ (26.5× higher)
Energy per correct answer (BM25)
4,145 J
BM25 vs MIRIX at 144,629 J (34.9× higher)
Agent Memory evaluates ten systems on MemoryAgentBench LongMemEval_S_* and shows BM25 at 47.0% accuracy with 582 kJ total energy, while Letta reaches 27.7% accuracy at 15,429 kJ. This MAIN_RESULT proves that system design choices in Agent Memory’s taxonomy cause over 47× differences in lifecycle energy per correct answer and orders-of-magnitude latency variation.
BENCHMARK
Agent Memory evaluates ten systems on MemoryAgentBench LongMemEval_S_* and shows BM25 at 47.0% accuracy with 582 kJ total energy, while Letta reaches 27.7% accuracy at 15,429 kJ. This MAIN_RESULT proves that system design choices in Agent Memory’s taxonomy cause over 47× differences in lifecycle energy per correct answer and orders-of-magnitude latency variation.
BENCHMARK
Energy per correct answer (J) for representative agent memory systems on LongMemEval_S_* with Qwen3-32B.
KEY INSIGHT
Agent Memory shows that construction energy for LLM-mediated memory systems can exceed query-phase energy across 300 queries, with energy per correct answer varying by over 47×.
This is counterintuitive because many builders assume serving latency and context size dominate costs, but Agent Memory reveals that background memory construction is often the true lifecycle bottleneck.
WHY IT MATTERS
Agent Memory gives practitioners a quantitative lens to choose between long-context, flat RAG, structure-augmented, and agentic memory designs based on energy, latency, and footprint, not just accuracy.
Builders can now design stateful agents that schedule construction, select appropriate paradigms, and manage fleet-scale memory stores with explicit tradeoffs between freshness, latency, and lifecycle cost.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
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.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…