Here is the translated markdown content:

Sharing an awesome list on LLM memory.

https://t.co/VRxMOviLa0

XiaomingX/awesome-ai-memory

Source: https://github.com/XiaomingX/awesome-ai-memory

Awesome AI Memory - Curated List of Long-Term Memory Projects for Large Language Models

A curated list focused on long-term memory (LTM) implementations for Large Language Models (LLMs), covering the complete ecosystem from underlying retrieval/storage paradigms, memory layer systems, and agent memory tools, to multimodal consistency, training, and alignment.

Core Updates (2026 Q2–Q3)

• RL-ified Memory Management: Projects like AgeMem transform the CRUD (Create, Read, Update, Delete) operations of long-term/short-term memory into trainable tool actions for agents, using three-stage RL (step-wise GRPO) for end-to-end optimization, replacing heuristic controllers.
• From “Atomic Facts” to Multi-Granularity: TriMem simultaneously maintains three granularities: raw dialogue segments, extracted facts, and synthesized portraits. It uses TextGrad to iteratively optimize prompts, enabling parameter-free lifelong evolution.
• Conflict-Aware and Fact Revision: MOSAIC, Infini Memory address the “facts change over time” problem by designing maintainable memory (topic documents / conflict detection), directly tackling the quality control challenge in the write path.
• Evaluation Shifts from “Recall” to “Action”: Benchmarks like MemoryAgentBench (ICLR’26), MemoryArena / Memoria-Bench / AMA-Bench (ICML’26), HaluMem, and LongMemEval-V2 push evaluation towards operational-level hallucinations, multi-session agentic tasks, and selective forgetting.
• Test-Time Memory Layering: Titans-as-a-Layer packages the neural long-term memory into a plug-and-play adapter (MAL), enabling the injection of long-term memory into multimodal models without altering the backbone.
• Local-First, Cross-Tool MCP Memory: Projects like Memorix, OKF-standard mcp-memory, implemented with SQLite / Knowledge Graph + Markdown, emphasize privatization and cross-host (Claude Code / Cursor / Codex) sharing.

Recent Major Advances (June–September 2026)

New Papers and Methods

New Datasets and Benchmarks

Complete list see docs/memory-evaluation-benchmarks.md.

New Trends

1. RL-ified Memory Management: Memory operations are no longer fixed pipelines but part of the policy, driven by reward signals to decide “what to remember and when to forget.”
2. Multi-granularity + Conflict-Aware Write Path: Shifting from “extracting atomic facts” to simultaneously preserving original text, facts, and portraits, while explicitly handling factual conflicts and revisions.
3. Operational-level / Agentic Evaluation: Evaluation moves from end-to-end QA scores to detailed stages like extraction, update, and forgetting, and shifts towards multi-session, actionable tasks.
4. Test-Time Memory Layering: Neural long-term memory evolves from “modifying architecture” to “pluggable layers,” lowering the barrier for injecting long-term memory into multimodal models.
5. Local-First and Standardization: MCP memory services emphasize local running and cross-host sharing, with emerging memory file format standards like OKF.
6. From “Remembering” to “Acting”: Benchmarks like Mem2ActBench and MemoryArena directly test the benefit of memory for tool calls and decision-making.

Technical Evolution of LLM Memory

To understand the positioning of the projects below, first clarify the development trajectory of memory technologies. Key verified milestones are listed chronologically (years/institutions per original papers).

Paradigm Foundations

• RAG (Retrieval-Augmented Generation, Meta/FAIR, 2020, arXiv 2005.11401): The retrieval + generation paradigm, considered foundational for LLM memory beyond parameters. Placing “knowledge” in external corpora and retrieving it as needed for injection into context.
• Generative Agents (Stanford, Park et al., 2023, arXiv 2304.03442): A town of 25 AI agents, proposing the “memory stream + retrieval + reflection” trifecta, a classic prototype for agent memory design.

Memory Mechanism Exploration (2023)

• MemoryBank (Zhong et al., 2023, arXiv 2305.10250, AAAI 2024): Inspired by the Ebbinghaus forgetting curve, applies decay and reinforcement to memory, accompanied by the LoCoMo long-range dialogue memory benchmark.
• LongMem (Wang et al., 2023, arXiv 2306.07174, NeurIPS 2023): Freezes the backbone LLM as a memory encoder, coupled with a Residual SideNet and cache memory bank, enabling long-term memory.
• MemGPT / Letta (UC Berkeley, Packer et al., 2023, arXiv 2310.08560): Applies OS virtual memory concepts to LLMs, stratifying context into primary context/external storage, using tool calls for “virtual context management.” The team later founded Letta to continue the open-source project.

Memory Compression and Episodic Memory (2024)

• ReadAgent (Google DeepMind, 2024): Simulates human reading, paginating long text and using gist memory to compress key points, an episodic memory framework.
• GraphRAG (Microsoft, 2024, arXiv 2404.16130): Extracts entity relationships from unstructured text to build knowledge graphs, with community summarization, enhancing retrieval-augmented generation for complex/global QA.
• HippoRAG (OSU-NLP-Group & Stanford et al., 2024, arXiv 2405.14831, NeurIPS 2024): Inspired by the hippocampal indexing theory, synergizes LLMs, knowledge graphs, and personalized PageRank for multi-hop retrieval akin to long-term memory.
• Memory3 / Explicit Memory (BAAI et al., 2024, arXiv 2407.01178): Externalizes knowledge as a sparse attention KV memory module, paralleling parametric and working memory as the “third type of memory.”

Production-Grade Memory Layers (2025)

• Mem0 (Mem0 Team, 2025, arXiv 2504.19413): A scalable long-term memory layer for production-grade AI agents, including vector Mem0 and graph Mem0g, claiming superior performance over OpenAI’s memory solutions in LLM-as-Judge evaluations while saving tokens.
• A-MEM (Xu et al., 2025, arXiv 2502.12110): Inspired by Luhmann’s Zettelkasten note-card method, uses dynamic indexing/linking for agent self-evolving structured memory, similar to but independent of the Zep/Graphiti path.

Architectural Long-Term Memory (2024-2025)

• Titans (Google, 2024-12, arXiv 2412.01427, NeurIPS 2025): Proposes a “neural long-term memory” module, driven by “surprise,” supporting test-time training/learning, combining RNN-style memory with attention for ultra-long contexts.

Learnable Memory Management and Evaluation Systems (2026)

• AgeMem / Agentic Memory (2026, arXiv 2601.01885, ACL 2026): Treats LTM/STM operations as tool actions for agent autonomous decision-making, trained with progressive RL, marking memory management’s shift from heuristic rules to learnable strategies.
• TriMem (2026, arXiv 2605.19952): Coexists three granularities: “raw dialogue + atomic facts + synthesized portraits” instead of single fact extraction, using TextGrad prompt optimization for parameter-free lifelong evolution.
• Infini Memory (2026, arXiv 2606.10677): Restructures memory into maintainable “topic documents,” coupled with iterative agentic retrieval, emphasizing evidence aggregation and fact revision.
• Titans-as-a-Layer (2026, arXiv 2606.08573): Packages test-time neural memory as a pluggable layer (MAL), shifting from “modifying architecture” to “adding modules,” first systematically extending to audio multimodality.
• Evaluation Systems (2026): MemoryAgentBench (ICLR), MemoryArena / Memoria-Bench / AMA-Bench (ICML), HaluMem, LongMemEval-V2, etc., refine evaluation to operational stages like extraction/update/forgetting and multi-session agentic tasks.

Evolution Main Line: External retrieval (RAG) → Agent memory streams and reflection (Generative Agents / MemGPT) → Memory compression and graphification (ReadAgent / GraphRAG / HippoRAG / Memory3) → Production-grade memory layers (Mem0 / A-MEM / Zep) → Architectural neural long-term memory (Titans) → Learnable memory management + operational evaluation (AgeMem / TriMem / MemoryAgentBench). Memory is evolving from “stuffing materials into prompts” to “the model itself learning to remember and managing what to remember.”

1. Integrated Memory Layers

These projects offer complex logical memory management, supporting multi-user, multi-session, automatic summarization, and knowledge extraction.

2. Agentic & Local Tools

Suitable for individual developers, single-machine agents, or integration into specific office workflows.

3. AI Frameworks with Memory Support

General AI development frameworks deeply integrated with long-term memory management capabilities.

4. Infrastructure & Storage

Foundational databases providing physical support for the above memory layers.

5. Practical Integration

1. Start Small: Beginners are advised to start with vector databases. Avoid introducing complex graph structures prematurely unless business needs demand it (e.g., complex plotline reasoning).
2. Long-Form Storytelling:
   • Prioritize using Mem0 or NovelGenerator to track character arcs.
   • Strategy: Generate and store chapter summaries first; retrieve summaries before generating new chapters to ensure coherence.
3. Ecosystem Compatibility:
   • For users of personal assistants like OpenClaw (formerly Clawdbot), consider using Mem0 or Letta, which offer mature SDKs.
4. Privacy-Sensitive: Local agents should prioritize Basic Memory or AgentCortex.

6. Research & Architecture

Advanced research focused on enhancing LLM memory capabilities at the training stage or architectural level.

7. MCP & Assistant Skills

Using the Model Context Protocol (MCP) or specific tool calls (Skills) to inject persistent memory capabilities into models.

8. Multimodal Consistency & Memory

Specialized tools and algorithms for maintaining character appearance, art style, and voice timbre consistency when generating images and videos.

9. Distributed Training

Ensuring absolute synchronization of parameters and gradients across multiple GPUs/nodes during large-scale model training.

10. Multi-Agent Coordination

Ensuring consistency of shared state, task progress, and contextual memory during agent team collaboration.

11. Online Training

Preventing catastrophic quality degradation and value drift during real-time user interaction or online fine-tuning (RLHF).

12. Continual Learning

Ensuring LLMs do not suffer catastrophic forgetting when adapting to new scenarios/domains, maintaining long-term memory continuity.

13. Fishbone Diagram of Memory Technology Evolution

Below is a fishbone diagram (cause-effect skeleton) outlining several main lines of “where LLM long-term memory capabilities come from,” to facilitate quick global comprehension. The fish head is “LLM Long-Term Memory Capability,” each “large bone” on the skeleton is a technical dimension, and branches are representative works.

┌─ RAG (2020, Meta) ................ Foundation of parametric-external memory ├─ GraphRAG (2024, MS) ........... Graph + community summarization Retrieval/External Memory ──────┼─ HippoRAG (2024) ............... Hippocampal indexing + PPR └─ Memory3 (2024, BAAI) .......... Explicit/sparse memory module │ ├─ Generative Agents (2023) . Memory stream + reflection ├─ MemGPT / Letta (2023) .... Virtual context stratification Mechanism/Architectural Memory ─────────────┼─ LongMem (2023) ............ Frozen encoder + cache bank ├─ ReadAgent (2024) ......... Pagination + gist compression ├─ Titans (2024-25, Google) . Neural LTM + test-time learning └─ Titans-as-a-Layer (2026) . Memory layering, plug-and-play │ ├─ MemoryBank (2023) ........ Forgetting curve decay Production-Grade Memory Layers ──────────────┼─ Mem0 / Mem0g (2025) ...... Vector + graph memory layer ├─ Zep / Graphiti ........... Temporal knowledge graph └─ A-MEM (2025) ............. Zettelkasten self-evolution │ ├─ AgeMem (2026) ............ RL-learned memory management strategy Learnable Memory Management (2026) ─────┼─ TriMem (2026) ............. Multi-granularity + parameter-free evolution ├─ Infini Memory (2026) ..... Topic documents + fact revision └─ MOSAIC (2026) ............ Conflict-aware structured memory │ ├─ MemoryAgentBench (ICLR'26) Four-dimensional capability decoupling Evaluation Systems (2026) ───────────┼─ MemoryArena / Memoria-Bench Multi-session agentic ├─ HaluMem ................... Operational-level hallucination └─ LongMemEval-V2 ........... Web agent environmental experience