2026 Memory Literature Scan¶

Last Updated: 2026-04-15

Methodology¶

This scan searched arxiv (cs.CL, cs.AI), Google Scholar, and major leaderboards (LongMemEval, LoCoMo) for 2026 and 2025 papers on LLM agent memory architectures.

Search queries: - "LLM memory" + "agent memory" + "long-term memory" (2026 filter) - "episodic memory LLM", "memory-augmented neural networks" - Forward citations from anchor papers (Hindsight 2512.12818, Supermemory, MemOS) - Leaderboard paper trails (LongMemEval, LoCoMo new entries) - Survey papers on memory systems ("Memory in the Age of AI Agents", etc.) - Graph-based memory, continual learning, test-time training

Coverage: cs.CL/cs.AI arxiv, OpenReview (ICLR 2026, NeurIPS 2025), specialized blogs (Zep, Emergentmind).

Tier 1 — 2026 Papers (Deep-dive Candidates)¶

1. Memory in the Age of AI Agents (Survey)¶

arxiv/link: 2512.13564
Authors/affil: Yuyang Hu et al. (46 authors, multi-institutional)
One-line: Consolidated taxonomy of agent memory across forms (token/parametric/latent), functions (factual/experiential/working), and dynamics (formation/evolution/retrieval).
Why notable: Establishes conceptual clarity for fragmented memory research. First unified framework treating memory as first-class agent primitive. Positions memory as distinct from RAG/context engineering.
Benchmarks reported: Analyzed 4+ recent benchmarks (LongMemEval, LoCoMo implicitly referenced); identifies persistent gaps in multi-session reasoning.
Relation to covered work: Organizes product/paper landscape into factual vs. experiential memory; subsumes Hindsight, Supermemory, MemOS conceptually.
Recommend deep dive: YES — This is the canonical 2026 taxonomy paper. Essential for framework coherence before diving into specific architectures.

2. Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers (Survey)¶

arxiv/link: 2603.07670
Authors/affil: Multi-authored survey (2026-02 or later)
One-line: Structured account of memory design/implementation/evaluation in LLM agents (2022–early 2026), proposing 3D taxonomy (temporal scope, representational substrate, control policy) and examining five mechanism families.
Why notable: Five mechanism families: (1) context-resident compression, (2) retrieval-augmented stores, (3) reflective self-improvement, (4) hierarchical virtual context, (5) policy-learned management. Shifts evaluation from static recall to multi-session tests interleaving memory with decision-making.
Benchmarks reported: References 4 recent multi-session agentic benchmarks; reports persisting limitations across systems.
Relation to covered work: Directly addresses product systems (memory modules in agents). Hierarchy aligns with Hindsight's tier 2 operations (retain/recall/reflect).
Recommend deep dive: YES — Complementary mechanisms view to Memory in Age of Agents; clarifies design space.

3. Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects¶

arxiv/link: 2512.12818
Authors/affil: Chris Latimer et al. (Vectorize.io, Washington Post, Virginia Tech)
One-line: Structured memory architecture organizing agent knowledge into four networks (world facts, experiences, entity summaries, beliefs) with retain/recall/reflect operations; implements Tempr (temporal entity memory) + Cara (reasoning with coherent opinion network).
Why notable: Addresses key gap: current systems blur evidence/inference, struggle long-horizon organization, offer limited reasoning transparency. Hindsight's tiered networks + opinion tracking is novel. Strongest empirical result: 83.6% accuracy on LongMemEval (vs. 39% full-context baseline) using 20B model, outperforming GPT-4o full-context.
Benchmarks reported: LongMemEval: 83.6% vs 39% baseline. LoCoMo results also reported.
Relation to covered work: Direct product ancestor (covered in engineering survey). This is the flagship academic paper establishing structured memory as SOTA.
Recommend deep dive: YES — Tier 1 anchor. Core reference for retain/recall/reflect paradigm.

4. Graph-based Agent Memory: Taxonomy, Techniques, and Applications¶

arxiv/link: 2602.05665
Authors/affil: Multi-author survey (Feb 2026)
One-line: Comprehensive graph-memory taxonomy covering short/long-term, knowledge/experience, non-structural/structural; evaluates knowledge graphs, temporal graphs, hypergraphs, hierarchical trees, hybrid graphs.
Why notable: Graph-based memory emerged as 2025–2026 frontier. First systematic taxonomy. Shows why graphs model relational dependencies, hierarchy, and efficient retrieval better than flat stores.
Benchmarks reported: Implicit comparison across graph variants; cites LoCoMo/LongMemEval but focuses on architectural properties.
Relation to covered work: Organizes graph papers (MAGMA, Zep/Graphiti, LiCoMemory, HyperMem, etc.). Subsumes product graph approaches (Graphiti from Zep).
Recommend deep dive: MAYBE — Excellent architectural survey. Recommend if exploring graph-specific deep-dives; may overlap with Memory in Age of Agents mechanistically.

5. MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents¶

arxiv/link: 2601.03236
Authors/affil: Multi-author (Jan 2026)
One-line: Extends single-graph KGs to multi-graph: orthogonal semantic, temporal, causal, entity graphs. Formulates retrieval as policy-guided traversal over relational views.
Why notable: Novel multi-graph framing unifies scattered context via orthogonal semantic/temporal/causal/entity relations. Outperforms state-of-the-art agentic memory on long-horizon LoCoMo/LongMemEval tasks.
Benchmarks reported: LoCoMo, LongMemEval (SOTA reported, specific numbers not extracted in search).
Relation to covered work: Advances Graphiti/Zep (single temporal KG) to richer relational model. Complements Hindsight's tier-2 networks (causality, entity linking explicit).
Recommend deep dive: YES — Multi-graph paradigm is novel and well-motivated. Strong empirical results.

6. SimpleMem: Efficient Lifelong Memory for LLM Agents¶

arxiv/link: 2601.02553
Authors/affil: Multi-author (Jan 5, 2026)
One-line: Three-stage pipeline: (1) semantic structured compression (distill unstructured to compact indexed units), (2) online semantic synthesis (intra-session redundancy elimination), (3) intent-aware retrieval (dynamic scope planning).
Why notable: Addresses token efficiency (64% improvement on LoCoMo; 30-fold inference token reduction). Emphasizes lossless compression + semantic synthesis. F1 improvement: 26.4% over baselines.
Benchmarks reported: LoCoMo: 64% improvement (as of Feb 2026 update); outperforms Claude-Mem. Achieves superior accuracy/efficiency/cost balance.
Relation to covered work: Orthogonal to graph-based; complements Hindsight's retain phase with compression. Text-centric (no graphs).
Recommend deep dive: YES — Token efficiency + multimodal support differentiates. Strong practical results.

7. LiCoMemory: Lightweight and Cognitive Agentic Memory for Efficient Long-Term Reasoning¶

arxiv/link: 2511.01448
Authors/affil: Multi-author (Nov 2025 submission, 2026 work)
One-line: CogniGraph—lightweight hierarchical graph using entities/relations as semantic indexing. Temporal + hierarchy-aware search with reranking; explicit hyperlinks to dialogue evidence.
Why notable: 23% accuracy improvement on LoCoMo/LongMemEval. Reframes KG as semantic index (not static repo). Reduces update latency; demonstrates efficient retrieval at scale.
Benchmarks reported: LoCoMo, LongMemEval (23% improvement over second-best). Temporal reasoning, multi-session consistency, retrieval efficiency all improve.
Relation to covered work: Intermediate between SimpleMem (lightweight) and MAGMA (multi-graph). Adds temporal awareness; lighter than MAGMA.
Recommend deep dive: YES — Practical efficiency gains + semantic indexing novel. Bridges graph/compression approaches.

8. Zep: A Temporal Knowledge Graph Architecture for Agent Memory¶

arxiv/link: 2501.13956 (published Jan 2025, 2026 refinements documented)
Authors/affil: Preston Rasmussen et al. (Zep company)
One-line: Graphiti engine—temporally-aware KG dynamically synthesizing unstructured (conversational) + structured (business) data; maintains fact validity windows (non-lossy history).
Why notable: Production-grade temporal fact management (facts have validity windows; old facts invalidated not deleted). Outperforms MemGPT on Deep Memory Retrieval (94.8% vs 93.4%). P95 latency 300ms; used in production CRM/compliance/healthcare.
Benchmarks reported: Deep Memory Retrieval (DMR): 94.8% vs MemGPT 93.4%.
Relation to covered work: Forward-cited anchor from plan. Graphiti is embedded in production (Mem0, etc.). This is the canonical temporal KG paper.
Recommend deep dive: YES — Production maturity + temporal validity windows novel. Validates graph-based approach at scale.

9. Adaptive Memory Admission Control for LLM Agents (A-MAC)¶

arxiv/link: 2603.04549
Authors/affil: Guilin Zhang et al. (published ICLR 2026 Workshop MemAgent, Mar 4, 2026)
One-line: Decomposes memory admission as structured decision: five interpretable factors (future utility, factual confidence, semantic novelty, temporal recency, content type prior) evaluated pre-storage.
Why notable: Elevates memory admission to first-class control (vs. opaque LLM judgment or heuristics). Explicit tradeoffs between coverage/reliability/efficiency. Novel control-theoretic framing.
Benchmarks reported: Not explicit in abstract; ICLR workshop paper likely empirical on LoCoMo/LongMemEval derivatives.
Relation to covered work: Orthogonal to storage architecture (works with any backend). Complements Hindsight's retain phase with structured gating.
Recommend deep dive: MAYBE — Strong control-theory contribution. Skip if focusing on retrieval; prioritize if studying memory lifecycle (retain → admit → recall → reflect).

10. A-MEM: Agentic Memory for LLM Agents (NeurIPS 2025, extended 2026)¶

arxiv/link: 2502.12110
Authors/affil: Wujiang Xu et al. (NeurIPS 2025 poster)
One-line: Zettelkasten-inspired agentic memory: dynamic indexing/linking creates interconnected knowledge networks. New memories generate structured notes (descriptions, keywords, tags); trigger contextual updates to historical memories.
Why notable: Zettelkasten formalism (bidirectional links, atomic notes) applied to LLM memory. Continuous refinement of understanding as new memories arrive. Empirical SOTA on 6 foundation models.
Benchmarks reported: 6 foundation models tested; reported SOTA improvements (specific benchmarks unclear from abstract).
Relation to covered work: Orthogonal organization principle (Zettelkasten) vs. graphs/compression. Complements graph approaches; more fine-grained note-linking.
Recommend deep dive: MAYBE — Novel organizational metaphor. Recommend if deep-diving organizational principles; may overlap with graph-based approaches functionally.

Tier 2 — 2025 Papers¶

1. Memoria: A Scalable Agentic Memory Framework for Personalized Conversational AI¶

arxiv/link: 2512.12686 (Dec 14, 2025 submission)
Authors/affil: Multi-author
One-line: Modular hybrid framework integrating dynamic session summarization + weighted KG-based user modeling. Four modules: conversation logging, user modeling, session summarization, context-aware retrieval.
Why notable: Hybrid short-term (summarization) + long-term (KG user model) architecture. Achieves 87.1% benchmark accuracy, 38.7% latency reduction, token-efficient. Plug-and-play design.
Benchmarks reported: 87.1% accuracy (benchmark unspecified); 38.7% latency reduction, reduced token usage.
Relation to covered work: Adds personalization (user trait/preference modeling) vs. generic architectures. Similar to Hindsight beliefs network but KG-structured.
Recommend deep dive: MAYBE — Personalization angle novel. Skip if focusing on core long-term reasoning; prioritize if user modeling is in scope.

2. MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems¶

arxiv/link: 2510.17281 (Oct 2025 submission)
Authors/affil: Multi-author (THUIR, Hugging Face dataset)
One-line: Comprehensive benchmark (declarative + procedural memory) with explicit/implicit user feedback. Tests continual learning, knowledge retention, task adaptation across domains/languages/input-output lengths.
Why notable: First benchmark providing all memory types + feedback data. Shows existing systems weak on procedural knowledge utilization and continual learning efficiency. Reveals memory systems not ready for real-world feedback loops.
Benchmarks reported: Open-sourced at Hugging Face. Baseline results on RAG + SOTA memory systems.
Relation to covered work: Benchmarking (not architecture). Complements LongMemEval/LoCoMo by testing continual learning.
Recommend deep dive: MAYBE — Evaluation framework. Prioritize if studying benchmark design; otherwise secondary.

3. Test-Time Training for Long-Context LLMs (TTT-E2E & Query-Only TTT)¶

arxiv/link: 2512.13898 (Dec 2025, "Let's (not) just put things in Context") + 2512.23675 (End-to-End TTT)
Authors/affil: Multi-author (MIT, industry)
One-line: Two variants: TTT-E2E (compress context to weights via next-token prediction, O(1) inference latency) and Query-Only TTT (lightweight query-matrix adaptation, reuse KV cache).
Why notable: Novel test-time learning paradigm. TTT-E2E achieves 2.7x speedup (128K context), 35x (2M context). Avoids external memory entirely; learning is built-in.
Benchmarks reported: Loss curves on long-context tasks; inference latency comparisons (vs. full attention, Mamba 2).
Relation to covered work: Orthogonal to external memory (embedded learning). Complements retrieval-based systems; reduces memory need.
Recommend deep dive: MAYBE — Represents different memory paradigm (learning-based vs. storage-based). Recommend if studying internal adaptation; secondary if external memory is focus.

4. Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations¶

arxiv/link: 2602.19320 (Feb 22, 2026)
Authors/affil: Dongming Jiang et al.
One-line: Introduces concise taxonomy of Memory-Augmented Generation (4 structures: Lightweight Semantic, Entity-Centric/Personalized, Episodic/Reflective, Structured/Hierarchical). Empirical analysis of benchmark saturation, metric misalignment, backbone variance, latency overhead.
Why notable: Critical analysis revealing fragile empirical foundations: benchmarks underscaled, metrics misaligned with semantic utility, performance backbone-dependent, system overhead overlooked. Identifies key pain points.
Benchmarks reported: Analyzes (doesn't report raw scores) LongMemEval, LoCoMo, others. Shows metric sensitivity to backbone (e.g., GPT-3.5 vs GPT-4 wildly different).
Relation to covered work: Meta-analysis of Tier 1 systems. Identifies evaluation pitfalls across Hindsight, SimpleMem, graph-based systems.
Recommend deep dive: YES — Critical read for methodology soundness. Highlights what's NOT robust in current memory systems.

5. AgeMem: Learning Unified Long-Term and Short-Term Memory Management for LLM Agents¶

arxiv/link: 2601.01885 (Jan 5, 2026)
Authors/affil: Yi Yu et al.
One-line: Unified framework integrating LTM + STM as agent policy. Memory operations (store/retrieve/update/summarize/discard) exposed as tool actions. RL training via three-stage progressive GRPO.
Why notable: Moves from separate LTM/STM (heuristic-driven) to learned unified policy. Autonomously decides what/when to store. Superior task performance, higher-quality LTM, efficient context on 5 long-horizon benchmarks.
Benchmarks reported: 5 long-horizon benchmarks (specific names/scores not extracted, but reported improvements across all).
Relation to covered work: Complements structured architectures (Hindsight, Memoria) with learned control. Similar intent to A-MAC but uses RL.
Recommend deep dive: MAYBE — Learning-based control novel. Recommend if studying policy-driven memory; may be secondary to structural architectures.

Tier 3 — Pre-2025 Reference Only¶

Paper	Year	Core Idea	Absorbed By
HippoRAG	2024	Neurobiologically-inspired RAG; Personalized PageRank over entity graphs for multi-hop retrieval.	HippoRAG 2 (2026); Zep/Graphiti; MAGMA builds on entity graph framing
MemGPT	2023	Hierarchical memory (core context, buffer, archival); external storage with learned read/write.	All modern systems; benchmark standard (Deep Memory Retrieval).
MemoryBank	2023	Ebbinghaus forgetting curve applied to agent memory decay.	SimpleMem (semantic compression); Hindsight (reflection handles decay).
RAG (Retrieval-Augmented Generation)	2020	External store + semantic retrieval.	Foundation for all Tier 1 retrieval systems; subsumed by graph/compression variants.
Transformer XL, Compressive Transformers	2018–2019	In-context compression via segment-level recurrence.	Basis for hierarchical context approaches; superseded by external memory.
Neural Turing Machines, Differentiable Neural Computers	2014–2016	Learned addressable external memory; attention-based retrieval.	Precursor framing to modern agentic memory; theoretical foundation (Turing-complete + memory = universality).

Gaps I Couldn't Fill¶

Specific numerical results on LoCoMo/LongMemEval for MAGMA, HyperMem: Papers reference SOTA but search-accessible abstracts didn't yield exact accuracy percentages. (Full PDFs needed.)
2026 papers on multimodal memory + agents: Only found episodic memory benchmark (2501.13121). Likely gap in search coverage.
Memory consolidation / sleep-like processes in LLMs: Possible 2026 work; didn't surface in arxiv searches. May exist in neuroscience/biology venue crossover.
Certified/trustworthy memory: Only tangential mention in Anatomy (backbone-dependent results). Likely pre-2026 or absent.
Memory for multi-agent systems: One mention (Awesome-Efficient-Agents repo), but no dedicated 2026 paper found. May be in workflow/multi-turn literature.
Hybrid internal + external memory: TTT-E2E (internal learning) and external storage are largely separate literatures. Unified frameworks may be emerging post-April 2026.

Summary¶

Tier 1 (2026) top 5 for deep dive: 1. Memory in the Age of AI Agents (2512.13564) — canonical taxonomy 2. Hindsight (2512.12818) — flagship structured memory + best empirical results 3. MAGMA (2601.03236) — multi-graph innovation 4. SimpleMem (2601.02553) — token efficiency + compression 5. LiCoMemory (2511.01448) — practical graph-based efficiency

Supporting Tier 2 (2025): - Anatomy of Agentic Memory (2602.19320) — critical evaluation meta-analysis - MemoryBench (2510.17281) — continual learning benchmark - AgeMem (2601.01885) — learned unified memory control

Leaderboard papers to track: - LongMemEval/LoCoMo leaderboards for late April–summer 2026 entries - ICLR 2026 MemAgents workshop proceedings (submitted papers, accepted posters)