AuthorsPreston Rasmussen, Pavlo Paliychuk, Travis Beauvais et al.
TL;DR
Zep uses the Graphiti temporal knowledge graph with episodic, semantic, and community memory to reach 71.2% on LongMemEval vs 60.2% full-context while cutting latency by ~90%.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LongMemEval shows LLM performance drops even when full 115k token conversations are stuffed into the context window.
This hurts enterprise chat assistants that must track cross-session memory and temporal reasoning, where naive full-context RAG becomes slow, costly, and unreliable.
HOW IT WORKS
Zep centers on Episode Subgraph, Semantic Entity Subgraph, Community Subgraph, and a Search–Reranker–Constructor pipeline inside the Graphiti temporal knowledge graph.
You can think of Zep like a brain with episodic memory for events, semantic memory for facts, and a card catalog that clusters related concepts into communities.
This design lets Zep maintain bi-temporal edges, invalidate outdated facts, and retrieve compact, time-aware context that plain context windows and static RAG cannot express.
DIAGRAM
This diagram shows how Zep processes a query through search, reranking, and construction over Graphiti's temporal knowledge graph.
DIAGRAM
This diagram shows how Zep ingests datasets, builds Graphiti, and runs memory benchmarks against full-context baselines.
PROCESS
Knowledge Graph Construction
Zep ingests Episodes into the Episode Subgraph, extracts entities and facts into the Semantic Entity Subgraph, and clusters them into the Community Subgraph using Graphiti.
Search
Given a query, Zep applies cosine semantic similarity, BM25 full text, and breadth first search over Graphiti to collect candidate edges, entities, and communities.
Reranker
Zep applies Reciprocal Rank Fusion, Maximal Marginal Relevance, and cross encoder rerankers to prioritize candidates and emphasize frequently mentioned or nearby nodes.
Constructor
Zep's constructor formats selected facts with tvalid and tinvalid plus entity and community summaries into a compact context string for the LLM agent.
KEY CONTRIBUTIONS
Temporally aware Graphiti knowledge graph
Zep introduces Graphiti with Episode Subgraph, Semantic Entity Subgraph, and Community Subgraph, tracking tvalid, tinvalid, t′created, and t′expired for each fact edge.
State of the art memory benchmarks
Zep reaches 94.8% on Deep Memory Retrieval with gpt-4-turbo and 98.2% with gpt-4o-mini, slightly exceeding MemGPT and full conversation baselines.
LongMemEval accuracy and latency gains
On LongMemEval, Zep with gpt-4o improves accuracy from 60.2% to 71.2% and cuts average latency from 28.9 s to 2.58 s using ~1.6k tokens.
RESULTS
DMR Score gpt 4 turbo
94.8%
+1.4 over MemGPT
DMR Score gpt 4o mini
98.2%
+0.2 over Full conversation
LongMemEval Score gpt 4o
71.2%
+11.0 over Full context
LongMemEval Latency gpt 4o
2.58 s
vs 28.9 s full context
Deep Memory Retrieval tests multi session chat recall, while LongMemEval stresses long term interactive memory with ~115k token conversations. Zep shows that temporal graph memory can beat full context baselines in both accuracy and latency.
BENCHMARK
Deep Memory Retrieval tests multi session chat recall, while LongMemEval stresses long term interactive memory with ~115k token conversations. Zep shows that temporal graph memory can beat full context baselines in both accuracy and latency.
BENCHMARK
Accuracy on Deep Memory Retrieval and LongMemEval comparing Zep to key baselines.
KEY INSIGHT
On LongMemEval, Zep with gpt-4o uses only about 1.6k context tokens yet scores 71.2% vs 60.2% for 115k token full context.
This is surprising because many assume feeding the entire conversation to an LLM is always best, but Zep shows structured temporal memory can be both shorter and more accurate.
WHY IT MATTERS
Zep unlocks scalable, temporally aware memory where agents can reason over years of conversations and updates without ballooning context windows.
Builders can now deploy chat systems that maintain cross session preferences and temporal facts with low latency, instead of relying on brittle summaries or expensive full history prompts.
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…