Shalini Goyal maps 10 memory systems for AI agents

Shalini Goyal published a multi-post overview of 10 memory systems AI engineers should understand for building agents, framing memory as a core determinant of agent capability. In her view, systems without memory behave more like advanced autocomplete than dependable assistants, and in many practical settings strong retrieval matters more than simply scaling the model. The thread is notable because it ties the gap between one-off answer tools and durable agent systems to memory-layer design.

Confirmed

• Goyal splits the foundation into short-term and long-term memory. Short-term memory acts like an agent’s RAM, holding the current conversation, recent actions, immediate context, and temporary information; long-term memory stores user preferences, conversation history, project information, and persistent cross-session knowledge.
• Semantic memory stores facts and concepts, making it useful for knowledge assistants, education systems, and research tools. Episodic memory stores past interactions, completed tasks, user behavior patterns, and workflow history so an agent can learn from experience.
• Vector memory represents information as embeddings for similarity search, semantic retrieval, RAG, and knowledge systems. Graph memory stores entities, relations, and context, which Goyal says is better suited to multi-hop reasoning, richer context understanding, and more intelligent retrieval.
• Procedural memory stores instructions, workflows, and agent behaviors—how things should be done—making it important for autonomous agents that must handle multi-step business processes. Retrieval memory covers the pipeline of storing information, fetching relevant context, ranking results, and generating an answer.
• Shared memory lets multiple agents exchange context, coordinate actions, and stay consistent. Hybrid memory combines several memory types—rather than relying on one architecture alone—to build more capable agents.

Why it matters

• Goyal argues modern AI applications need to preserve context across sessions, retrieve relevant information on demand, learn from interactions, and improve over time. In that framing, the memory layer is what turns an answer engine into a real system. In a follow-up post, eyishazyer echoed that assessment and said agents are unlikely to become reliable systems without memory.

2026-07-27 ~ 2026-07-27 · 12 related posts

• [source] 10 AI memory systems every agent engineer should know — goyalshaliniuk · 2026-07-27
• AI agents need short-term memory to keep working context alive — goyalshaliniuk · 2026-07-27
• AI memory systems: semantic memory stores facts, long-term memory preserves context — goyalshaliniuk · 2026-07-27
• Short-term and long-term memory are the base layer of AI assistants — goyalshaliniuk · 2026-07-27
• Episodic and semantic memory split experience from factual knowledge — goyalshaliniuk · 2026-07-27
• Vector memory and episodic memory give agents retrieval plus experience — goyalshaliniuk · 2026-07-27
• Graph memory and vector memory power modern AI retrieval stacks — goyalshaliniuk · 2026-07-27
• [source] Good retrieval may matter more than a bigger model for AI agents — goyalshaliniuk · 2026-07-27
• Procedural and retrieval memory are becoming core agent primitives — goyalshaliniuk · 2026-07-27
• Shared memory lets multi-agent systems coordinate and stay consistent — goyalshaliniuk · 2026-07-27
• [source] Hybrid and shared memory systems aim to make multi-agent AI consistent — goyalshaliniuk · 2026-07-27
• AI agents need memory to become real systems, says new roundup — eyishazyer · 2026-07-27