AuthorsShu Wang, Edwin Yu, Oscar Love et al.
TL;DR
MemMachine uses ground-truth episodic storage plus contextualized retrieval to reach 93.0% on LongMemEvalS and 0.9169 on LoCoMo with gpt-4.1-mini.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Standard RAG and context windows struggle with multi-session interactions, leading to brittle personalization and factual drift over long horizons.
On LoCoMo, the OpenAI baseline only reaches an overall score of 0.5290, limiting reliable personalization, temporal reasoning, and multi-hop conversational recall.
HOW IT WORKS
MemMachine centers on Short-term memory, Long-term memory, Profile memory, and a Retrieval Agent to store raw episodes and retrieve sentence-level evidence.
Think of MemMachine like RAM plus disk plus a card catalog: short-term memory is RAM, long-term episodic storage is disk, and contextualized retrieval is the catalog that pulls neighboring pages.
This design lets MemMachine preserve exact conversational ground truth while selectively surfacing relevant episode clusters, something a plain context window or naive RAG cannot achieve.
DIAGRAM
This diagram shows how MemMachine processes a query through short-term search, long-term vector search, contextualization, and reranking before returning episodes.
DIAGRAM
This diagram shows how MemMachine varies ingestion and retrieval settings across LongMemEvalS configurations to measure accuracy gains.
PROCESS
Data Ingestion
MemMachine converts each message into an Episode with producer, timestamp, session id, and custom metadata, then dispatches it to Short-term memory and Long-term memory.
Sentence Extraction
MemMachine segments each Episode into sentences using NLTK Punkt, links them back to episodes, and embeds them into the vector-backed Long-term memory.
Contextualized Retrieval
MemMachine runs vector search to find nucleus episodes, expands them with neighboring episodes into clusters, and reranks clusters before assembling STM and LTM context.
Retrieval Agent
MemMachine optionally routes the query through the Retrieval Agent, choosing direct search, SplitQuery, or ChainOfQuery, then passes retrieved context to the answer LLM.
KEY CONTRIBUTIONS
Ground-truth-preserving architecture
MemMachine stores raw conversational episodes and indexes at sentence level, avoiding per-message extraction and enabling about 80% input token reduction versus Mem0 on LoCoMo.
Contextualized retrieval
MemMachine introduces contextualized retrieval that expands nucleus matches with neighboring episode context, then reranks clusters, improving multi-hop and temporal reasoning accuracy.
Retrieval Agent for multi-hop reasoning
MemMachine adds the Retrieval Agent with ToolSelectAgent, SplitQuery, and ChainOfQuery, reaching 93.2% on HotpotQA hard and 92.6% on WikiMultiHop under randomized noise.
RESULTS
LoCoMo overall score (gpt-4.1-mini)
91.69%
+13.79 over OpenAI baseline
LongMemEvalS ablation best
93.0%
+7.0 over MemMachine baseline C5
HotpotQA hard Retrieval Agent accuracy
93.2%
+2.0 over MemMachine declarative search
Mem0 comparison input tokens
~80% less
MemMachine vs Mem0 on LoCoMo memory mode
On LoCoMo, which tests very long-term conversational memory, MemMachine reaches 91.69% with gpt-4.1-mini while using about 80% fewer input tokens than Mem0. On LongMemEvalS, which probes extraction, temporal reasoning, updates, and multi-session reasoning, MemMachine’s best configuration reaches 93.0%, showing that ground-truth episodic storage plus tuned retrieval yields strong long-horizon recall.
BENCHMARK
On LoCoMo, which tests very long-term conversational memory, MemMachine reaches 91.69% with gpt-4.1-mini while using about 80% fewer input tokens than Mem0. On LongMemEvalS, which probes extraction, temporal reasoning, updates, and multi-session reasoning, MemMachine’s best configuration reaches 93.0%, showing that ground-truth episodic storage plus tuned retrieval yields strong long-horizon recall.
BENCHMARK
LLM Judge Score on LoCoMo (overall).
BENCHMARK
Overall LLM score on LongMemEvalS across key MemMachine configurations.
KEY INSIGHT
MemMachine finds that GPT-5-mini beats GPT-5 by 2.6 percentage points on LongMemEvalS when paired with the Edwin3 prompt and tuned retrieval.
This is counterintuitive because larger models are usually assumed strictly better, but MemMachine shows prompt–model co-optimization can make a smaller, cheaper model superior.
WHY IT MATTERS
MemMachine unlocks cost-efficient, ground-truth-preserving long-term memory with contextualized retrieval and multi-hop-aware routing that scale across sessions and benchmarks.
Builders can now deploy personalized agents that remember exact past interactions, handle complex multi-hop questions, and stay within tight token budgets using smaller LLMs like GPT-5-mini.
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.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…