Agent Memory Atlas

Agent Memory Atlas is a release-candidate field guide for builders who need more than another paper list. It gives humans an oriented map of AI-agent memory systems and gives agents structured source records, taxonomy axes, tags, confidence scores, provenance, and verification hints.

Who This Is For

Persona	What you get
Agent builders	A fast map from architecture choices to source material, benchmarks, and implementation references.
Researchers	A categorized starting set with explicit uncertainty instead of hidden citation quality problems.
Agentic tools	JSONL, CSV, indexes, and stable IDs for retrieval, ranking, and planning workflows.

Map

Start Here

ID	Source	Category	Status	Why it matters
src-001	MemGPT: Towards LLMs as Operating Systems (Packer et al., Oct 2023)	Foundational Papers	public candidate	OS-inspired tiered memory architecture
src-002	Generative Agents (Park et al., Apr 2023)	Foundational Papers	public candidate	Memory stream + reflection; ablation showed emergent coordination requires consolidation
src-003	A-MEM: Agentic Memory for LLM Agents (Xu et al., Feb 2025)	Foundational Papers	public candidate	Zettelkasten-inspired note-linking; atomic units with dynamic connections
src-007	CoALA (Sumers et al., Sep 2023)	Foundational Papers	public candidate	Four-type cognitive memory model
src-009	Zep/Graphiti (Zep team, Jan 2025)	Foundational Papers	public candidate	Bi-temporal graph knowledge base
src-016	Memory in the Age of AI Agents (Zhang et al., Dec 2025)	Survey Papers	public candidate	Agent memory survey and unification framework
src-021	LoCoMo (Maharana et al., ACL 2024)	Benchmark Papers	public candidate	Long-conversation memory benchmark
src-029	Mem0	Open-Source Systems	public candidate	Personalized memory layer for LLM applications
src-038	A-MEM	Open-Source Systems	public candidate	Zettelkasten note-linking implementation
src-047	OpenAI Agents SDK	Agent Frameworks	public candidate	Tracing built-in

Why This Is Different

A plain paper list	Agent Memory Atlas
Optimized for browsing titles	Optimized for deciding what to build, test, or cite next
Mostly human-readable Markdown	Human field guide plus machine-readable JSONL, CSV, and indexes
Categories hide uncertainty	Every record carries verification and URL status
Hard for agents to consume reliably	Stable IDs, tags, taxonomy axes, confidence scores, provenance, and verification hints

Current Preview Status

• Total source records: 122
• Public candidate URLs: 93
• Needs claim check: 0
• Reference-only entries: 29
• Needs URL fix before public release: 0
• Needs public sanitization: 0

This repository is still private pending maintainer approval. The public-facing truth surface is claim-safe: public candidate records avoid quantitative/popularity claims, unresolved citations are quarantined as reference_only, and unresolved claim records are not allowed in release-candidate output.

Taxonomy Axes

Axis	Sources	Description
Memory Architecture	19	How memory is represented, tiered, linked, and persisted.
Parametric Memory	9	How systems store durable knowledge in weights, adapters, edits, and continual-learning mechanisms.
Working Memory	7	How agents manage active context, attention, KV cache, and task-local reasoning state.
Retrieval & Access	5	How agents find and prioritize relevant memories.
Lifecycle & Governance	8	How memories are written, consolidated, updated, forgotten, and audited.
Benchmarks & Evaluation	15	Datasets and methods for testing long-horizon memory behavior.
Systems & Implementation	43	Open-source systems, stores, frameworks, and protocols used to build memory layers.
Human Field Guide	16	Surveys, essays, and community resources that orient builders quickly.

Agent-Readable Files

File	Purpose
data/sources.jsonl	Canonical source records, one JSON object per source
data/sources.csv	Spreadsheet-friendly source table
data/taxonomy_matrix.csv	Counts by taxonomy axis and category
data/benchmark_matrix.csv	Citation-only benchmark/evaluation matrix
taxonomy.json	Machine-readable taxonomy source of truth
indexes/source_index.json	Stable ID lookup for retrieval and linking
indexes/tag_index.json	Tag-to-source lookup
indexes/claim_index.json	Should be empty for release-candidate output
atlas.manifest.json	Export metadata and release gate status
atlas.agent-map.json	Suggested agent workflows over the atlas

Source Field Guide

Agent Frameworks

ID	Source	Priority	Status	Evidence	Why it matters
src-041	LangGraph	P0	public candidate	no numeric claim	Reducer patterns
src-042	CrewAI	P0	public candidate	no numeric claim	Memory coordination
src-043	AutoGen	P1	public candidate	no numeric claim	Teachable agents
src-044	LlamaIndex	P1	public candidate	no numeric claim	RAG integration
src-045	Pydantic AI	P1	public candidate	no numeric claim	Type-safe interfaces
src-046	Mastra	P2	public candidate	no numeric claim	Observation memory
src-047	OpenAI Agents SDK	P1	public candidate	no numeric claim	Tracing built-in
src-048	IBM BeeAI / ACP	P1	public candidate	no numeric claim	ACP protocol

Benchmark Papers

ID	Source	Priority	Status	Evidence	Why it matters
src-021	LoCoMo (Maharana et al., ACL 2024)	P0	public candidate	no numeric claim	Long-conversation memory benchmark
src-022	LongMemEval (Wu et al., ICLR 2025)	P0	public candidate	no numeric claim	Long-context memory evaluation benchmark
src-023	BEAM (Tavakoli et al., ICLR 2026)	P0	reference only	no numeric claim	Extreme-scale memory benchmark concept
src-024	MemoryArena (He et al., ICML 2026)	P0	reference only	no numeric claim	Action-memory coupling
src-025	MemoryAgentBench (Hu et al., ICLR 2026)	P0	public candidate	no numeric claim	Cognitive memory competency benchmark
src-026	LifeBench (2026)	P1	reference only	no numeric claim	Non-declarative eval
src-027	LoCoMo-Plus (Li et al., 2026)	P1	reference only	no numeric claim	Constraint-consistency evaluation beyond surface matching
src-028	MemBench (Tan et al., ACL 2025)	P2	reference only	no numeric claim	Multi-metric

Benchmarks and Evaluation

ID	Source	Priority	Status	Evidence	Why it matters
src-117	HippoRAG: Neurobiologically Inspired Long-Term Memory for LLMs	P1	public candidate	no numeric claim	Long-term memory retrieval reference
src-118	RoleLLM: Benchmarking Role-Playing Abilities of LLMs	P1	public candidate	no numeric claim	Role-playing evaluation reference
src-119	LearnAct: Few-Shot Mobile GUI Agent Benchmark	P1	public candidate	no numeric claim	Mobile GUI agent benchmark
src-120	RepairAgent: Autonomous LLM-Based Program Repair	P1	public candidate	no numeric claim	Program-repair agent benchmark
src-121	ToolLLM: Real-World API Tool Use Benchmark	P1	public candidate	no numeric claim	Tool-use evaluation reference
src-122	AgentEvolver: Self-Evolving Agent System	P2	public candidate	no numeric claim	Self-evolving agent evaluation reference
src-123	WebWeaver: Dynamic Outlines for Deep Research	P2	public candidate	no numeric claim	Deep-research evidence structuring reference

Community Resources

ID	Source	Priority	Status	Evidence	Why it matters
src-098	Agent-Memory-Paper-List (Shichun-Liu)	P0	public candidate	no numeric claim	Agent memory paper list
src-099	Awesome-Memory-for-Agents (Tsinghua C3I)	P0	public candidate	no numeric claim	Agent memory benchmark collection
src-100	Awesome-Agent-Memory-Papers (yyyujintang)	P1	public candidate	no numeric claim	Agent memory paper dashboard

Conferences and Workshops

ID	Source	Priority	Status	Evidence	Why it matters
src-095	ICLR 2026 MemAgents Workshop	P0	reference only	no numeric claim	Memory workshop
src-096	NeurIPS 2025: AI Assistants in the Wild	P1	reference only	no numeric claim	Edge focus
src-097	L2M2 @ ACL 2025: LLM Memorization	P1	reference only	no numeric claim	Privacy risks

Context Engineering

ID	Source	Priority	Status	Evidence	Why it matters
src-080	LongContext vs. RAG: When to Use Which (2024)	P1	reference only	no numeric claim	Systematic comparison of context windows vs retrieval
src-081	Lost in the Middle (Liu et al., 2023)	P1	public candidate	no numeric claim	U-shaped attention; models attend to start/end
src-082	Chain-of-Thought Prompting (Wei et al., NeurIPS 2022)	P2	public candidate	no numeric claim	Foundational context structuring technique
src-083	HyDE: Zero-Shot Dense Retrieval (2022)	P1	public candidate	no numeric claim	Hypothetical document embeddings improve retrieval
src-084	RAG-Fusion (2024)	P2	public candidate	no numeric claim	Multi-query + RRF for improved retrieval

Databases and Storage

ID	Source	Priority	Status	Evidence	Why it matters
src-049	pgvector (PostgreSQL)	P0	public candidate	no numeric claim	PostgreSQL vector extension for retrieval storage
src-050	Qdrant	P0	public candidate	no numeric claim	Self-hosted value
src-051	Neo4j	P0	public candidate	no numeric claim	Graph + vector unified
src-052	Weaviate	P1	public candidate	no numeric claim	Hybrid search
src-053	Milvus / Zilliz Cloud	P1	public candidate	no numeric claim	Distributed vector database for large indexes
src-054	KuzuDB (community forks)	P1	public candidate	no numeric claim	Embedded graph
src-055	Redis	P1	public candidate	no numeric claim	Semantic cache
src-056	LanceDB	P2	public candidate	no numeric claim	Embedded vector
src-057	SQLite	P1	public candidate	no numeric claim	Verbatim thesis
src-058	FalkorDB	P2	public candidate	no numeric claim	GraphRAG

Foundational Papers

ID	Source	Priority	Status	Evidence	Why it matters
src-001	MemGPT: Towards LLMs as Operating Systems (Packer et al., Oct 2023)	P0	public candidate	no numeric claim	OS-inspired tiered memory architecture
src-002	Generative Agents (Park et al., Apr 2023)	P0	public candidate	no numeric claim	Memory stream + reflection; ablation showed emergent coordination requires consolidation
src-003	A-MEM: Agentic Memory for LLM Agents (Xu et al., Feb 2025)	P0	public candidate	no numeric claim	Zettelkasten-inspired note-linking; atomic units with dynamic connections
src-004	Memory OS of AI Agent (Kang et al., May 2025)	P0	public candidate	no numeric claim	Hierarchical storage and recall architecture
src-005	PISA (Jia et al., Oct 2025)	P1	public candidate	no numeric claim	Schema taxonomy
src-006	Hindsight (Latimer et al., Dec 2025)	P0	public candidate	no numeric claim	TEMPR retrieval
src-007	CoALA (Sumers et al., Sep 2023)	P0	public candidate	no numeric claim	Four-type cognitive memory model
src-008	Mem0 (Mem0 team, Apr 2025)	P0	public candidate	no numeric claim	LLM fact extraction for memory write paths
src-009	Zep/Graphiti (Zep team, Jan 2025)	P0	public candidate	no numeric claim	Bi-temporal graph knowledge base
src-010	Graph-based Agent Memory (Yang et al., Feb 2026)	P1	public candidate	no numeric claim	Six cognitive memory types including sentiment; graph lifecycle taxonomy
src-011	The AI Hippocampus (Jia et al., Jan 2026)	P1	public candidate	no numeric claim	Implicit/explicit/agentic
src-012	From Storage to Experience (Luo et al., 2026)	P1	reference only	no numeric claim	Evolutionary stages: storage to reflection to experience
src-013	Cognitive Memory in LLMs (Anonymous, Apr 2025)	P2	public candidate	no numeric claim	Psychology-to-AI memory type mapping
src-014	DAMCS: Decentralized Adaptive KG Memory (AAAI 2025)	P1	public candidate	no numeric claim	Shared knowledge graph memory architecture
src-015	G-Memory (NeurIPS 2025)	P1	reference only	no numeric claim	Hierarchical graph memory concept

MCP and Tooling

ID	Source	Priority	Status	Evidence	Why it matters
src-059	MCP Official Specification	P0	reference only	no numeric claim	Official Model Context Protocol specification
src-060	@modelcontextprotocol/server-memory	P0	public candidate	no numeric claim	Official server
src-061	mcp-memory-service (doobidoo)	P1	public candidate	no numeric claim	Hybrid search MCP
src-062	ai-memory-mcp (alphaonedev)	P1	public candidate	no numeric claim	Tiered MCP
src-063	Mem0 MCP (cloud-hosted)	P1	reference only	no numeric claim	Cloud MCP model
src-064	Cognee MCP	P1	public candidate	no numeric claim	KG MCP
src-065	OWASP MCP Top 10	P0	reference only	no numeric claim	MCP security
src-066	MCP Security Research (Adversa AI)	P0	public candidate	no numeric claim	MCP threats

Memory Lifecycle Research

ID	Source	Priority	Status	Evidence	Why it matters
src-067	FSFM: Framework for Selective Forgetting (2026)	P0	reference only	no numeric claim	Selective forgetting framework
src-068	FadeMem: Dual-Layer Differential Decay (2026)	P0	reference only	no numeric claim	Dual-layer memory decay framework
src-069	AMV-L: Adaptive Memory Value Lifecycle (2026)	P0	reference only	no numeric claim	Adaptive memory-value lifecycle framework
src-070	Memoria: Git for Agent Memory (2026)	P1	public candidate	no numeric claim	Snapshot, branch, merge, rollback; CoW in Rust
src-071	Focus Agent: Intra-Trajectory Compression (Verma, 2026)	P1	reference only	no numeric claim	Intra-trajectory memory compression framework
src-072	Neotoma: Deterministic State Layer (2026)	P1	public candidate	no numeric claim	Versioned timeline
src-073	ZenBrain: 7-Layer Neuroscience Architecture (2026)	P2	reference only	no numeric claim	Neuroscience-inspired memory consolidation framework
src-074	MuninnDB: Cognitive Database (2026)	P2	public candidate	no numeric claim	ACT-R + Hebbian + Bayesian at engine level

Multi-Agent Memory

ID	Source	Priority	Status	Evidence	Why it matters
src-076	G-Memory (NeurIPS 2025)	P0	reference only	no numeric claim	Organizational memory graph concept
src-077	Emergent Collective Memory in Decentralized MAS (2025)	P1	reference only	no numeric claim	Stigmergic collective memory concept
src-078	Multi-Agent Memory from Computer Architecture (2026)	P1	reference only	no numeric claim	Architecture perspective
src-079	Collaborative Memory: Multi-User Sharing (2025)	P2	reference only	no numeric claim	Dynamic access control

Open-Source Systems

ID	Source	Priority	Status	Evidence	Why it matters
src-029	Mem0	P0	public candidate	no numeric claim	Personalized memory layer for LLM applications
src-030	Graphiti	P0	public candidate	no numeric claim	Temporal KG
src-031	Letta	P0	public candidate	no numeric claim	Self-editing memory
src-032	Cognee	P0	public candidate	no numeric claim	Poly-store knowledge graph memory system
src-033	Hindsight	P0	public candidate	no numeric claim	Temporal episodic memory retrieval system
src-034	LangMem	P1	public candidate	no numeric claim	Procedural memory
src-035	OMEGA	P1	public candidate	no numeric claim	Local-first
src-036	Memoria	P1	public candidate	no numeric claim	Version control
src-037	MuninnDB	P1	public candidate	no numeric claim	Cognitive primitives
src-038	A-MEM	P1	public candidate	no numeric claim	Zettelkasten note-linking implementation
src-039	True Memory	P2	public candidate	no numeric claim	Single-file SQLite memory system
src-040	Engram AI	P2	public candidate	no numeric claim	Neuroscience model

Parametric Memory

ID	Source	Priority	Status	Evidence	Why it matters
src-101	ROME: Locating and Editing Factual Associations in GPT	P0	public candidate	no numeric claim	Model editing method for factual memory
src-102	MEMIT: Mass-Editing Memory in a Transformer	P0	public candidate	no numeric claim	Batch model editing method
src-103	AlphaEdit: Null-Space Constrained Knowledge Editing	P1	public candidate	no numeric claim	Knowledge editing method
src-104	WISE: Rethinking Knowledge Memory for Lifelong Model Editing	P1	public candidate	no numeric claim	Lifelong model editing method
src-105	ELDER: Enhancing Lifelong Model Editing with Mixture-of-LoRA	P1	public candidate	no numeric claim	LoRA-based lifelong editing method
src-106	K-Adapter: Infusing Knowledge into Pre-Trained Models	P1	public candidate	no numeric claim	Adapter-based knowledge infusion
src-107	MemLoRA: Distilling Expert Adapters for On-Device Memory Systems	P1	public candidate	no numeric claim	Adapter-based memory system
src-108	Character-LLM: A Trainable Agent for Role-Playing	P2	public candidate	no numeric claim	Trainable role memory for agents
src-109	ELLA: An Efficient Lifelong Learning Algorithm	P2	public candidate	no numeric claim	Continual-learning baseline

Product Documentation

ID	Source	Priority	Status	Evidence	Why it matters
src-085	Mem0 Documentation	P1	reference only	no numeric claim	Config patterns
src-086	Graphiti Documentation	P1	reference only	no numeric claim	Temporal API
src-087	Letta Documentation	P1	reference only	no numeric claim	Self-edit API
src-088	Cognee Documentation	P1	reference only	no numeric claim	Retrieval modes
src-089	Hindsight Documentation	P1	reference only	no numeric claim	TEMPR

Survey Papers

ID	Source	Priority	Status	Evidence	Why it matters
src-016	Memory in the Age of AI Agents (Zhang et al., Dec 2025)	P0	public candidate	no numeric claim	Agent memory survey and unification framework
src-017	Anatomy of Agentic Memory (Jiang et al., Feb 2026)	P0	public candidate	no numeric claim	MAG taxonomy
src-018	Memory for Autonomous LLM Agents (Du, Mar 2026)	P0	public candidate	no numeric claim	Control policy
src-019	Rethinking Memory Mechanisms (Huang et al., Feb 2026)	P1	public candidate	no numeric claim	Survey of memory mechanisms for agents
src-020	A Survey on Memory Mechanism of LLM-based Agents (Zhang et al., Apr 2024)	P2	public candidate	no numeric claim	ACM TOIS baseline

Technical Essays

ID	Source	Priority	Status	Evidence	Why it matters
src-090	"RAG Is on Life Support" (Latimer, Vectorize)	P1	public candidate	no numeric claim	Memory vs RAG thesis
src-091	"Agent Memory Engineering" (Bustamante, May 2026)	P1	public candidate	no numeric claim	Agent memory lifecycle patterns
src-092	"The Lethal Trifecta" (Willison, Jun 2025)	P0	public candidate	no numeric claim	MCP risks
src-093	"MemPalace Benchmarks: Debunked" (Vectorize, Apr 2026)	P1	public candidate	no numeric claim	Audit
src-094	"Storage Is Not Memory" (Sauron Labs, 2026)	P1	reference only	no numeric claim	Verbatim thesis

Working Memory

ID	Source	Priority	Status	Evidence	Why it matters
src-110	Agent S: An Open Agentic Framework That Uses Computers Like a Human	P0	public candidate	no numeric claim	Computer-use agent context management
src-111	DeepAgent: A General Reasoning Agent with Scalable Toolsets	P1	public candidate	no numeric claim	Tool-using agent context management
src-112	ACON: Optimizing Context Compression for Long-Horizon LLM Agents	P1	public candidate	no numeric claim	Long-horizon context compression
src-113	Sculptor: Active Context Management for LLM Agents	P1	public candidate	no numeric claim	Active context management
src-114	MemAgent: Multi-Conv RL-based Memory Agent	P1	public candidate	no numeric claim	Memory agent for long-context handling
src-115	Efficient Streaming Language Models with Attention Sinks	P1	public candidate	no numeric claim	Streaming attention method
src-116	RazorAttention: Efficient KV Cache Compression Through Retrieval Heads	P1	public candidate	no numeric claim	KV-cache compression method

Release Gate

Public visibility still requires:

• Run secret and private-infrastructure scans on the exact release commit.
• Get explicit maintainer approval before changing repository visibility.