10 Open-Source Repositories and Architecture Patterns Giving AI Agents Persistent Memory

As LLM-powered autonomous agents evolve from isolated single-turn Q&A toward multi-session and personalized workflows, persistent memory architectures that retain context across sessions have become a fundamental building block.

Image source: @DivyanshT91162 via X

This guide examines prominent open-source projects and architectural patterns that move beyond standard stateless API calls to help agents preserve user preferences, interaction history, and evolving factual knowledge.

Cross-Session Context and User Preferences: Mem0 & Zep

Retaining context across different conversations is essential for personalization without inflating token budgets on every single prompt.

• Mem0: Provides a persistent memory layer designed for AI applications and agents. It systematically organizes user preferences, past conversation history, and key facts across sessions, retrieving them dynamically when needed.
• Zep: Focuses on long-term memory infrastructure for AI agents, offering pipelines that quickly retrieve relevant historical context to maintain conversational continuity.

Both projects reduce context window bloat and manage token expenses while ensuring consistent, personalized agent responses over time.

Temporal Knowledge Tracking and Structuring: Graphiti & Cognee

Facts and entity relationships change over time. Graph-based architectures are emerging to represent these dynamic shifts accurately.

• Graphiti: Implements a temporal knowledge graph architecture that tracks how facts, entities, and events evolve. By linking information to timestamps, it allows agents to distinguish past states from updated current realities.
• Cognee: Converts raw unstructured data into structured knowledge engines that AI agents can navigate, link, and query systematically.

Stateful Autonomous Agents and Infrastructure: Letta & Memary

Frameworks that empower agents to actively read and modify their own memory layers are expanding rapidly.

• Letta (formerly MemGPT lineage): A stateful agent framework engineered so agents can directly manage their memory hierarchy and maintain continuous state across interaction boundaries.
• Memary: Provides an experimental memory infrastructure exploring how agents store, categorize, and recall information beyond individual chats.
• Memvid & SimpleMem: Investigate lightweight file-based portable memory formats and long-term memory compression techniques designed to minimize LLM context overhead.

Architecture Selection and Production Considerations

Deploying persistent memory layers into production workflows requires matching the data model and security posture to the application's needs:

• Storage Structure Variety: Implementations range from vector databases and key-value embeddings to temporal knowledge graphs (Graphiti) and file-based portable stores (Memvid).
• Latency and Data Hygiene: Teams must account for synchronization latency during memory writes and implement sanitization routines to prevent malicious prompt injections from polluting persistent stores.

Sources

• Divyansh Tiwari (@DivyanshT91162): 10 GitHub Repos That Give AI Agents a Memory
• Leonard Rodman (@RodmanAi): Open-Source Memory Repositories for AI Agents