PaperlayerView exact revision on arXiv

Source-linked AI summary

Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents

Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, Tara Khani

arXiv:2604.22085v1cs.AI

TL;DR

Persistent memory is a key architectural requirement for autonomous agents, while existing hybrid graph systems impose ingestion and retrieval overhead. Memanto combines typed semantic memory, conflict resolution, and no-indexing semantic search, achieving state-of-the-art accuracy with zero-cost ingestion and lower complexity.

  • Problem

    Persistent memory is needed as language-model agents move toward multi-step reasoning, tool use, and long-horizon execution, while production systems require high accuracy, low latency, cost efficiency, and reduced operational complexity.

  • Method

    Memanto combines Moorcheh’s no-indexing Information Theoretic Search with a 13-category typed semantic memory schema and automated conflict resolution.

  • Results

    Memanto achieves state-of-the-art accuracy on LONGMEMEVAL (89.8%) and LOCOMO (87.1%) using vector-only retrieval, zero-cost ingestion, and a single retrieval query.

  • Takeaways & Limitations

    The ablation study identifies retrieval recall, rather than architectural complexity, as the dominant performance driver for agentic memory.

  • Takeaways & Limitations

    Approximately 5% of LONGMEMEVAL and 6–7% of LOCOMO questions contain labeling inconsistencies, creating a benchmark-dependent performance ceiling.

Abstract

from arXiv · show

The transition from stateless language model inference to persistent, multi session autonomous agents has revealed memory to be a primary architectural bottleneck in the deployment of production grade agentic systems. Existing methodologies largely depend on hybrid semantic graph architectures, which impose substantial computational overhead during both ingestion and retrieval. These systems typically require large language model mediated entity extraction, explicit graph schema maintenance, and multi query retrieval pipelines. This paper introduces Memanto, a universal memory layer for agentic artificial intelligence that challenges the prevailing assumption that knowledge graph complexity is necessary to achieve high fidelity agent memory. Memanto integrates a typed semantic memory schema comprising thirteen predefined memory categories, an automated conflict resolution mechanism, and temporal versioning. These components are enabled by Moorcheh's Information Theoretic Search engine, a no indexing semantic database that provides deterministic retrieval within sub ninety millisecond latency while eliminating ingestion delay. Through systematic benchmarking on the LongMemEval and LoCoMo evaluation suites, Memanto achieves state of the art accuracy scores of 89.8 percent and 87.1 percent respectively. These results surpass all evaluated hybrid graph and vector based systems while requiring only a single retrieval query, incurring no ingestion cost, and maintaining substantially lower operational complexity. A five stage progressive ablation study is presented to quantify the contribution of each architectural component, followed by a discussion of the implications for scalable deployment of agentic memory systems.

I. INTRODUCTION

Persistent memory is a central engineering requirement for long-horizon agents, while existing systems often trade competitive accuracy for substantial architectural and operational complexity. Memanto addresses this tradeoff with typed semantic memory, conflict resolution, and simpler retrieval infrastructure.

  • Persistent state across sessions is a central challenge as language models become autonomous agents capable of multistep reasoning and tool use.
  • Production agent memory must combine high accuracy, low latency, cost efficiency, and reduced operational complexity amid accelerating agent adoption.Industry projections cited in the introduction estimate market growth from 7.8 billion dollars to more than 52 billion dollars by 2030.
  • Existing frameworks commonly combine knowledge graphs, temporal graph databases, multi-query retrieval, and LLM-driven ingestion, creating substantial computational and operational overhead.The paper names Mem0, Zep, Letta, and A-MEM as examples of this increasingly complex approach.
  • Memanto combines Moorcheh’s no-indexing Information Theoretic Search with structured memory typing and automated conflict resolution to avoid graph infrastructure and ingestion overhead.The architecture is presented as requiring a single retrieval query and no schema management.
  • Memanto’s contribution is framed as a production-grade vector architecture with zero-cost ingestion and a five-stage ablation study evaluating retrieval and inference choices.The paper also situates typed memory within episodic, semantic, and procedural distinctions from cognitive science and prior agent-memory work.

B. Memory Surveys and Taxonomies (2024 to 2026)

Recent surveys and systems organize agent memory through cognitive categories, taxonomies, and increasingly complex hybrid architectures. Prior work shows that graph and reflective mechanisms can improve functionality, but their overhead motivates simpler retrieval-centered designs.

  • Survey literature classifies memory systems by forms, functions, dynamics, and symbolic-versus-neural paradigms.
  • Architectural surveys position memory as a layer connecting perception, agent reasoning, learning, planning, and action across time.
  • Production agent memory has converged on hybrid systems combining dense vector representations with structured knowledge graphs.
  • MemGPT/Letta uses virtual-memory paging and recursive summarization, which can introduce latency variability and information loss during precise textual recall.
  • Mem0’s graph-augmented configuration adds synchronous entity extraction, embedding and index updates, and graph synchronization, turning low-latency writes into multisecond procedures.
  • Zep, A-MEM, and reflective frameworks add temporal indexing, insertion-time inference, or iterative retrieval and reflection, increasing latency, cost, or system complexity.
  • Simple retrieval-based systems have outperformed more elaborate memory hierarchies on existing benchmarks, challenging the necessity of architectural complexity.

D. The Indexing and Ingestion Bottleneck

Indexing and ingestion delays constrain memory systems that must support immediate retrieval during interactive agent trajectories. The surrounding literature therefore emphasizes fine-grained, high-recall retrieval and evaluation protocols that distinguish memory quality from language-model reasoning.

  • Traditional vector databases use approximate-nearest-neighbor indexes that create delays between data ingestion and query availability.
  • Immediate post-storage retrieval matters because indexing latency can impair an agent’s same-trajectory reasoning.
  • LONGMEMEVAL decomposes memory performance into indexing, retrieval, and reading, identifying granularity, key construction, query formulation, and reading strategy as important factors.
  • Long-context processing can degrade for information positioned in the middle of extended contexts, reinforcing targeted retrieval over raw context length.
  • High-recall retrieval with post-retrieval verification is presented as a robustness strategy, although hierarchical and graph-based alternatives add system complexity.
  • LONGMEMEVAL contains 500 questions across six memory categories, while LOCOMO evaluates long-form multisession dialogues involving single-hop, multihop, open-domain, and temporal reasoning.
  • As context windows grow, benchmark results may increasingly reflect language-model reasoning rather than memory architecture quality, motivating more targeted evaluation protocols.
  • Memanto’s design principles were informed by production requirements, benchmark failure modes, and a structured dialogue identifying passive context injection and related memory gaps.

A. Six Desiderata for Production Agentic Memory

Production agent memory is specified through six desiderata covering queryability, temporal awareness, provenance, typing, and related operational requirements. Memanto implements these requirements through a typed service architecture with automatic memory processing and semantic retrieval.

  • Queryable memory lets agents retrieve task-relevant information on demand instead of receiving a static context injection.Static injection can exceed the context window, include irrelevant content, or omit recently stored facts.
  • Temporal awareness requires temporal queries, versioning, and relevance-decay signals because memories differ in urgency over time.
  • Confidence and provenance metadata distinguish explicit facts, inferred patterns, and potentially outdated information for calibrated use of retrieved context.
  • Typed memory separates episodic, semantic, and procedural information so each category can be stored and queried with appropriate semantics.
  • Memanto exposes remember, recall, and answer endpoints for automatic typing and conflict detection, ITS-powered retrieval, and retrieval-augmented generation.
  • The system’s production coverage is presented against six desiderata, with Memanto shown as covering all six while other systems have gaps across conflict detection, ingestion cost, and temporal or typed capabilities.

C. The Moorcheh Foundation: Information Theoretic Search

Moorcheh’s Information Theoretic Search foundation combines compressed, index-free semantic retrieval with typed memory and conflict handling. The resulting design targets efficient retrieval while preserving structured and coherent agent memory.

  • Information-Theoretic Search: 32× compression preserves retrieval-relevant signal without measurable loss, while the engine replaces conventional cosine similarity with an information-theoretic distance measure.The engine’s core innovations are Maximally Informative Binarization and Efficient Distance Metric.
  • Information-Theoretic Search: 64–74% NDCG@10 and 9.6 ms distance calculation latency were achieved on MAIR, compared with 37–86 ms for PGVector and Qdrant.The engine also sustained 2,000+ queries per second without accuracy degradation and delivered a 6.6× end-to-end speedup versus Pinecone plus Cohere reranking.
  • Information-Theoretic Search: Index-free retrieval enables instant write-to-search availability by eliminating index construction.This property follows from the combined retrieval innovations described for the Moorcheh engine.
  • Typed Memory: Memanto organizes memories into 13 semantic categories with distinct retrieval semantics, priority weighting, type-filtered retrieval, and implicit priority and decay signals.Typing supports focused queries for categories such as commitments and decisions without polluting the result set.
  • Conflict Resolution: Conflict resolution detects semantic contradictions within the same memory type and namespace, then offers supersede, retain, or annotate options.The mechanism is intended to address accumulated contradictory memories that can erode reasoning coherence.

F. Temporal Versioning

Memanto provides temporal query modes and non-destructive memory supersession for reconstructing historical and current states. Its namespace and session model preserves access to memories across session boundaries while supporting operational audit artifacts.

  • Temporal Query Modalities: Memanto supports as-of, changed-since, and current-only queries for historical reconstruction, incremental updates, and non-superseded ground-truth state.These modalities cover distinct temporal retrieval needs.
  • Temporal Versioning: Superseded memories remain stored and marked, enabling full temporal reconstruction and addressing knowledge-update evaluation in LongMemEval.The design is described as relevant to compliance-sensitive deployments in regulated industries.
  • Namespaces and Sessions: Each agent has an independent memory namespace, while six-hour default sessions group memories temporally without restricting cross-session retrieval.All memories in a namespace remain accessible across session boundaries.
  • Operational Artifacts: Automated daily artifacts include session summaries, contradiction reports, and interactive conflict-resolution prompts persisted as local Markdown files and optionally synchronized to Moorcheh.These artifacts provide human-readable audit trails and machine-queryable context.
  • Evaluation Context: The evaluation isolates architectural decisions through a five-stage progressive ablation study before reporting final per-category accuracy and operational comparisons.The study covers two established agentic memory benchmarks.

A. Benchmarks and Evaluation Protocol

Memanto is evaluated on LongMemEval and LoCoMo using controlled retrieval and prompt ablations. Results show that expanding retrieval recall produces the largest gain, while prompt optimization contributes comparatively little and broader recall improves performance despite added token cost.

  • Benchmarks: 500 LongMemEval questions span six categories and evaluate information extraction, multi-session reasoning, temporal reasoning, knowledge update, and abstention.The standard setting uses approximately 115K tokens across approximately 50 sessions, with Claude Sonnet 4 as judge.
  • Benchmarks: LoCoMo tests Single-Hop, Multi-Hop, Open Domain, and Temporal reasoning across dialogues reaching 35 sessions and 300 turns.The benchmark averages approximately 9K tokens per dialogue.
  • Evaluation Protocol: All experiments use Memanto’s vector-only architecture with Moorcheh ITS as the sole retrieval backend, while adapted Hindsight prompts address refusal and lexical-variation evaluation artifacts.This protocol aims to improve consistency across systems.
  • Stage 1: Naive Baseline: 56.6% on LONGMEMEVAL and 76.2% on LOCOMO establish the Stage 1 naive baseline with k=10, threshold 0.15, and Claude Sonnet 4.The baseline uses standard semantic search with a minimally parameterized RAG configuration.
  • Stage 2: Recall Expansion: 20.4 percentage points on LONGMEMEVAL and 6.6 percentage points on LOCOMO are gained by increasing retrieval from k=10 to 40 and relaxing the threshold to 0.10.This is the largest single improvement across the five ablation stages.
  • Stage 3: Prompt Optimization: 2.2 percentage points on LONGMEMEVAL and 0.1 percentage points on LOCOMO result from prompt optimization with retrieval settings unchanged.The paper characterizes these improvements as marginal.
  • Stage 4: Maximum Recall: 5.8 percentage points on LONGMEMEVAL and 3.4 percentage points on LOCOMO result from maximum-recall retrieval with dynamic limits up to 100 chunks and threshold 0.05.The associated finding is that broader retrieval budgets consistently outperform engineering for retrieval precision.
  • Retrieval Limit and Cost: Accuracy curves plateau above k=60, with a clear inflection at k=40, while the k=10→40 gain on LME outweighs the approximately fourfold token-cost increase.The figure presents accuracy and average tokens per query together.

C. Final Results by Category

Memanto’s Stage 5 category results vary across interaction settings, with strongest performance on single-session queries and lowest performance on multi-session queries.

  • 100.0% accuracy was achieved on the Single-session Assistant category.
  • 95.7% accuracy was achieved on the Single-session User category.
  • 81.2% accuracy was achieved on Multi-session queries, the lowest reported category performance.The passage attributes this lower performance to synthesizing information distributed across extended interaction histories.

D. Comparative Results

Memanto achieves the highest accuracy among vector-only systems on both benchmarks while using a simpler, lower-overhead architecture than competing approaches.

  • Memanto achieves the highest accuracy among all vector-only systems on both benchmarks.
  • 22.9 percentage points separate Memanto from Mem0 on LONGMEMEVAL, and 20.2 percentage points separate them on LOCOMO.
  • Hindsight achieves higher overall accuracy on both benchmarks at a complexity score of 4 out of 4.Its architecture requires dynamic multi-query retrieval and structured reflection passes.
  • Memanto achieves sub-10 ms ingestion and sub-90 ms retrieval using a single query.The comparison passage contrasts this with multi-second round trips for graph-traversal systems.
  • Memanto requires only the Moorcheh Vector DB and API key, without additional infrastructure to configure or operate.
  • At 10K daily memory operations, estimated daily costs are $0.50 for Memanto, $2.32 for Mem0-Graph, and $1.70 for Zep.The reported annual savings relative to Mem0-Graph are approximately $662 per agent.

V. DISCUSSION

The discussion argues that retrieval recall, semantic matching quality, and ingestion simplicity can outperform architectural complexity in agentic memory systems.

  • Memanto challenges the assumption that knowledge graphs are necessary for high-quality agentic memory.
  • Broader semantically relevant raw context lets the LLM perform in-context reasoning instead of relying on lossy, schema-dependent precomputed graph pathways.
  • Information-theoretically grounded exact semantic matching can reduce the marginal value of knowledge-graph structural overhead.
  • Eliminating LLM extraction at ingestion enables sub-second write-to-retrieval feedback loops.
  • +28.4 pp improvement on LongMemEval came from expanding k from 10 to 100, compared with +2.2 pp from prompt optimisation.

C. Conflict Resolution as a Production Necessity

Memanto treats conflict handling and deployment scope as important production considerations, while acknowledging limitations in benchmark coverage, model dependence, scale evaluation, and multi-agent sharing.

  • Conflict resolution: Contradictory memories are not systematically tested by either LongMemEval or LoCoMo.
  • Conflict resolution: Without explicit conflict detection, user corrections and evolving contexts can produce memory poisoning and increasingly incoherent agent behaviour.
  • Benchmark scope: Both benchmarks target conversational settings, leaving research agents, code generation, and multi-agent coordination untested.
  • Benchmark quality: Approximately 5% of LONGMEMEVAL and 6–7% of LOCOMO questions contain reported labelling inconsistencies.
  • Inference model dependence: Final results use Gemini 3, which contributes +4.8 pp on LongMemEval.
  • Scale evaluation: Large-scale memory benchmarks with thousands of concurrent agents remain future work despite Moorcheh validation at 10M+ documents and 2,000+ QPS.
  • Multi-agent memory sharing: Memanto’s namespace architecture currently isolates agent memories, while shared multi-agent memory remains under development.

APPENDIX

The appendix specifies the reproducibility resources, software configuration, hardware requirements, retrieval settings, and current limitation of manual memory-type assignment.

  • Reproducibility: Benchmark evaluations are reproducible using the described configuration, with code and evaluation scripts publicly available on GitHub and datasets hosted on Hugging Face.
  • Software configuration: Evaluations use Memanto 2.1.4 with the latest production Moorcheh SDK, Claude Sonnet 4 in Stages 1–4, and Gemini 3 in Stage 5.
  • Current limitation: Memory types are manually selected by users at write time from the 13-type schema, while automated rule-based assignment is planned for a future release.
  • Retrieval configuration: The final configuration uses ITS threshold gating with a 100-chunk retrieval budget, a 0.05 similarity threshold, and one retrieval query per question.
Loading 2604.22085v1…