AuthorsJiaqi Chen, Yongqin Zeng, Shaoshen Chen et al.
TL;DR
CoreMem uses Riemannian retrieval plus Fisher-guided discrete token distillation to boost LOCOMO Judge accuracy to 0.540 (+0.009 over NaiveRAG) under 8 GB VRAM.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CoreMem targets dialogue agents where a typical long-dialogue session with 1,500 context tokens and 50 daily queries costs about $22.50 per month per user.
Under these conditions, a competitive embedding model already uses ∼3 GB VRAM, leaving little room for memory indexing and compression, so agents either forget history or become economically unsustainable.
HOW IT WORKS
CoreMem centers on Riemannian retrieval, Fisher-guided discrete token distillation, Residual Metric Fusion, and an Edge-Cloud Hybrid Architecture to unify retrieval and compression via information geometry.
You can think of CoreMem like a smart RAM plus disk system where a geometric index finds the right pages and a Fisher-based compressor rewrites them into dense summaries before sending to the cloud.
This information-geometric design lets CoreMem surface tail memories, bound compression-induced KL divergence, and keep prompts 20–30% of their original size, which a plain context window cannot achieve.
DIAGRAM
This diagram shows how CoreMem processes a user query through edge-side retrieval and Fisher-guided compression before cloud LLM generation.
DIAGRAM
This diagram shows how CoreMem is evaluated from datasets through retrieval variants and compressors to Judge accuracy and latency metrics.
PROCESS
Riemannian Retrieval
CoreMem uses Riemannian retrieval to treat embeddings as points on a statistical manifold and compute Fisher-Rao based Mahalanobis distances with Woodbury acceleration.
Residual Metric Fusion
CoreMem applies Residual Metric Fusion, combining min max normalized cosine similarity and Riemannian scores with α = 0.5 to balance tail memories and bridge facts.
Fisher-Guided Discrete Token Distillation
CoreMem runs Fisher-guided discrete token distillation, using diagonal Fisher information traces and structural protection masks to drop low sensitivity tokens under budget B.
Edge-Cloud Hybrid Architecture
CoreMem's Edge-Cloud Hybrid Architecture sends only the distilled context to the cloud LLM, which generates the final answer conditioned on compressed long-term memory.
KEY CONTRIBUTIONS
Riemannian retrieval
CoreMem introduces Riemannian retrieval with a Fisher-Rao metric and O(Ndr) Woodbury acceleration, dynamically setting λ = 10·mean(σ²) to stabilize Σ⁻¹ while mitigating hubness.
Fisher-guided discrete token distillation
CoreMem derives Fisher-guided discrete token distillation, using diagonal Fisher traces and structural masks to achieve compression ratios up to 67.4% with conditional accuracy 0.9756.
Edge-cloud hybrid memory architecture
CoreMem designs an Edge-Cloud Hybrid Architecture that fits within 8 GB VRAM, keeping peak edge usage around 3.2 GB while reducing prompt tokens by 20–30% on LOCOMO.
RESULTS
Judge
0.540
+0.009 over NaiveRAG on LOCOMO MiniLM-L6
ROUGE-L
0.227
+0.012 vs NaiveRAG on LOCOMO MiniLM-L6
Hit@10
0.532
+0.017 over NaiveRAG on LOCOMO MiniLM-L6
Prompt tokens
2.09M
−0.0 96M vs NaiveRAG on LOCOMO MiniLM-L6
These metrics come from LOCOMO with MiniLM-L6, which tests Single hop, Multi hop, Temporal, Open domain, and Adversarial dialogue QA. The 0.540 Judge accuracy and +4.51 pp Open domain gain show that CoreMem retrieves and compresses long term memories more effectively than cosine based NaiveRAG under the same VRAM budget.
BENCHMARK
These metrics come from LOCOMO with MiniLM-L6, which tests Single hop, Multi hop, Temporal, Open domain, and Adversarial dialogue QA. The 0.540 Judge accuracy and +4.51 pp Open domain gain show that CoreMem retrieves and compresses long term memories more effectively than cosine based NaiveRAG under the same VRAM budget.
BENCHMARK
Judge accuracy on LOCOMO (MiniLM-L6) comparing CoreMem-Fusion against cosine-based memory systems.
BENCHMARK
Overall Judge accuracy on LongMemEval-S for different retrieval strategies with MiniLM-L6 embeddings.
KEY INSIGHT
Under extreme compression with budget 250, CoreMem-V3 achieves the highest ROUGE-L (0.1379) yet the lowest Judge accuracy (0.1979) among compressors.
This is surprising because higher lexical overlap usually implies better semantic correctness, but CoreMem shows that preserving surface tokens without causal connectors can still break reasoning.
WHY IT MATTERS
CoreMem unlocks long term personalized dialogue on 8 GB edge devices by geometrically aware retrieval and Fisher bounded compression that preserve crucial tail memories.
Builders can now deploy agents that remember months of interactions, keep cloud costs manageable, and still reason over temporal and open domain queries without server grade hardware.
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.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…