AuthorsFengrong Wan, Chengcan Wu, Ningtao Lyu
TL;DR
SodaMem uses an evidence-grounded temporal graph memory with FactEvents and connection-density fusion to reach 92.8% accuracy on LongMemEval-S at $0.00161 per question.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents assisting users over weeks must remember what is currently true, not merely what was once said in append-only logs.
Flat RAG diaries and Markdown logs leave currency, provenance, and ordered temporal reasoning unresolved, causing incoherent personal-memory QA over multi-session histories.
HOW IT WORKS
SodaMem centers on FactEvents, a temporal graph store, hybrid BM25–dense indexes, and a planner–reader loop to keep user memory coherent and citable.
You can think of SodaMem like a card catalog plus timeline: every fact is a card with timestamps and links, and retrieval walks the catalog instead of a single long notebook.
This evidence-grounded temporal graph lets SodaMem enforce supersession, soft time windows, and citation-backed answers that a plain context window or flat RAG store cannot provide.
DIAGRAM
This diagram shows how SodaMem performs multi-tunnel retrieval and connection-density fusion when answering a user question.
DIAGRAM
This diagram shows how SodaMem is evaluated on LongMemEval-S, including ingest, answering, and cost measurement.
PROCESS
IngestSession
SodaMem runs IngestSession over each dialogue session, extracting FactEvents with provenance spans and normalized modality and dates before writing facts and edges.
Store and maintain
SodaMem persists FactEvents in SQLite with hybrid BM25–dense indexes, maintaining SUPERSEDES, CONTRADICTS, and UPDATES edges plus optional dreaming and timeline resolution.
MultiTunnelRetrieve
SodaMem executes MultiTunnelRetrieve, using graph, BM25, and embedding tunnels with validity gates and soft time bonuses to build a fused evidence pool ranked by connection density.
Planner–Reader Loop
SodaMem runs a planner–reader loop where the planner calls memory tools to expand evidence and the reader composes an evidence-grounded answer with mandatory citations.
KEY CONTRIBUTIONS
Problem framing
SodaMem isolates currency, temporal structure, provenance, and association as failure modes of Markdown and flat RAG for long-horizon personal assistants, guiding the FactEvent schema and graph design.
System
SodaMem introduces an ingest–store–planner–reader pipeline with typed FactEvents, hybrid multi-signal retrieval, supersession semantics, and a timeline-resolution layer targeting temporal-reasoning errors.
Cost–accuracy evaluation
SodaMem reports 92.8% accuracy on LongMemEval-S at mean $0.00161 per question and compiles a cost–accuracy map showing SodaMem strictly dominates several higher-cost, lower-accuracy systems.
RESULTS
Accuracy
92.8%
+15.0 percentage points over MemOS (77.8%)
Questions
464/500 correct
best of N=3 runs on LongMemEval-S
Mean cost per 103Q
$1.61
Flash planner+reader spend excluding ingest and judge
Median cost per 103Q
$1.11
≈14.6k tokens per question, 25% cheaper than mean
On LongMemEval-S, which probes multi-session recall, updates, temporal reasoning, preference, and abstention, SodaMem reaches 92.8% accuracy at Flash-tier cost. This MAIN_RESULT shows SodaMem can maintain a coherent, updatable, and citable personal memory while staying near the accuracy frontier without Opus or GPT-4o-level spend.
BENCHMARK
On LongMemEval-S, which probes multi-session recall, updates, temporal reasoning, preference, and abstention, SodaMem reaches 92.8% accuracy at Flash-tier cost. This MAIN_RESULT shows SodaMem can maintain a coherent, updatable, and citable personal memory while staying near the accuracy frontier without Opus or GPT-4o-level spend.
BENCHMARK
Accuracy on LongMemEval-S and relative position of SodaMem among systems with estimable API cost.
KEY INSIGHT
SodaMem achieves 92.8% accuracy on LongMemEval-S at a mean cost of only $0.00161 per question with deepseek-v4-flash.
This is surprising because systems like agentmemory V4 and Mem0 2026 use Claude Opus and GPT-4o at roughly 10–40× higher estimated cost to gain only 1.6–3.4 percentage points of accuracy.
WHY IT MATTERS
SodaMem unlocks a practical way to keep long-horizon personal agents evidence-grounded, temporally coherent, and cost-efficient using FactEvents and connection-density fusion.
Builders can now deploy multi-week memory with explicit supersession, soft temporal reasoning, and auditable citations at Flash-tier spend, instead of relying on opaque long-context prompts or expensive high-end models.
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…