AuthorsSuozhao Ji, Baodong Wu, Zehao Wang et al.
TL;DR
Infini Memory uses topic-structured documents plus agentic retrieval to reach 64.7% overall on MemoryAgentBench, +19.2 points over HippoRAG-v2.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Infini Memory targets four recurring failures: memory fragmentation, memory conflict, compression loss, and insufficient retrieval, where evidence is scattered and contradictions remain unresolved.
These failures break long-horizon tasks like multi-hop reasoning and selective forgetting, leaving agents unable to reconcile updated facts or assemble temporally grounded evidence across sessions.
HOW IT WORKS
Infini Memory centers on topic documents, a CURRENT buffer, REWRITE_CURRENT, a topic library, and an agentic reader that together maintain and read persistent memory.
You can think of Infini Memory like a card catalog plus a staging tray: new notes pile in the tray, then get filed into well-labeled folders that a librarian can browse interactively.
This design lets Infini Memory revise facts, preserve temporal metadata, and iteratively expand local context in ways a plain context window or single-shot vector retrieval cannot.
DIAGRAM
This diagram shows how Infini Memory-A uses tool calls to iteratively search, inspect, and expand evidence from topic documents and the CURRENT buffer.
DIAGRAM
This diagram shows how Infini Memory is evaluated on MemoryAgentBench and how maintenance and retrieval ablations are configured.
PROCESS
Memory Extraction
Infini Memory extracts structured memory fragments from each interaction and appends them into the CURRENT buffer with sequence and temporal metadata.
Buffered Writing and Flush
Infini Memory monitors the CURRENT buffer and flushes it when token or time thresholds are met, preparing content for consolidation without rewriting the topic library each turn.
Rewrite CURRENT and Topic Routing
Infini Memory rewrites CURRENT into REWRITE_CURRENT, then uses topic routing to update or create topic documents inside the topic library, resolving local redundancy and contradictions.
Agentic Retrieval over Topic Library
Infini Memory-A uses the agentic reader and memory tools to search the topic library and CURRENT buffer, expanding local context and assembling evidence before answering.
KEY CONTRIBUTIONS
Document-Based Persistent Memory Architecture
Infini Memory introduces topic documents plus a topic library maintained via buffered writing and consolidation, achieving 81.2% Accurate Retrieval on MemoryAgentBench with gpt-5-mini.
Agentic Retrieval over Structured Memory Tools
Infini Memory-A adds an agentic reader that iteratively calls tools like search and read_lines over topic documents, improving overall MemoryAgentBench score to 64.7% versus 61.3% for Infini Memory-H.
Ablation Analysis of Maintenance and Retrieval
Infini Memory’s ablations show structural maintenance and retrieval both matter: disabling split and merge drops LongMemEval accuracy from 76.0% to 69.3%, while summary-only retrieval falls to 41.7%.
RESULTS
Overall Score
64.7%
+19.2 over HippoRAG-v2
Accurate Retrieval
81.2%
+12.5 over REMem
Test-Time Learning
51.0%
+4.4 over REMem
Long-Range Understanding
68.6%
+25.4 over REMem
On MemoryAgentBench, which tests Accurate Retrieval, Test-Time Learning, Long-Range Understanding, and Selective Forgetting, Infini Memory-A’s 64.7% overall score demonstrates that topic-structured documents plus agentic retrieval substantially improve long-term memory use for LLM agents.
BENCHMARK
On MemoryAgentBench, which tests Accurate Retrieval, Test-Time Learning, Long-Range Understanding, and Selective Forgetting, Infini Memory-A’s 64.7% overall score demonstrates that topic-structured documents plus agentic retrieval substantially improve long-term memory use for LLM agents.
BENCHMARK
Overall Score on MemoryAgentBench across memory systems using gpt-5-mini with 4096-token chunks.
BENCHMARK
Accuracy on LongMemEval (S⋆) for different Infini Memory maintenance and retrieval configurations.
KEY INSIGHT
Infini Memory’s ablation shows that removing structural split and merge drops LongMemEval accuracy by 6.7 points, while using summary-only retrieval collapses it by 37.6 points.
This is surprising because many systems emphasize sophisticated memory structures, yet Infini Memory reveals that how the reader interacts with simple topic documents can matter even more than the underlying storage complexity.
WHY IT MATTERS
Infini Memory unlocks long-term agents that maintain editable, interpretable topic documents while still achieving strong retrieval, learning, and forgetting behavior across sessions.
Builders can now deploy memory-heavy agents without specialized vector or graph databases, layering agentic retrieval tools over plain text stores to get maintainable, high-quality long-term memory.
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…