From Human Memory to Machine Memory: A Field Guide to AI Memory Architecture

Why a series about memory

For the last two years, almost every interesting paper about AI agents has had the word "memory" in its title. Vector databases. Episodic stores. Sleep-time agents. Reflection. Reconsolidation. The vocabulary keeps growing, and most of it is borrowed straight from cognitive psychology and neuroscience textbooks. That is not an accident. The brain has been working on the same problem (how to store, retrieve, and update useful information across time) for hundreds of millions of years. AI engineers are starting to notice that the answers are sitting in plain sight.

This series is a guided tour through the parallels. It maps 18 components of human memory to their closest analogs in modern AI systems and tells you, plainly, which ones are well-solved, which are still active research, and which are basically not built yet.

The four academic frameworks worth knowing

Four papers shaped the way people now talk about agent memory. You will see references to all of them throughout the series:

• CoALA (Sumers et al., 2023) borrowed Tulving's classification (working, episodic, semantic, procedural) and used it to lay out a clean taxonomy for language agents.
• Memory in the Age of AI Agents (Hu et al., 2025) extended that taxonomy with a Forms / Functions / Dynamics axis, arguing that the older "short-term vs long-term" split was no longer enough.
• From Human Memory to AI Memory (Wu et al., 2025) introduced a 3D-8Q taxonomy mapping memory across object, form, and time.
• Memp: Exploring Agent Procedural Memory (Fang et al., 2025) zoomed in on procedural memory, the least-built layer in agents today, and proposed a Build / Retrieve / Update lifecycle.

A fifth paper, HippoRAG (Gutierrez et al., NeurIPS 2024), deserves a special mention because it is the most biologically explicit system in production: it directly implements the hippocampal indexing theory using a knowledge graph as the index and an LLM as the parametric cortex. It outperforms standard RAG by up to 20% on multi-hop QA while being 10-30x cheaper.

The headline finding

Across all 18 subsystems, one pattern keeps repeating: retrieval and indexing are well-solved. Consolidation, forgetting, and metacognition are not. Modern HNSW indexes, transformer attention, and hybrid search give us mature analogs for hippocampal pattern completion and spreading activation. But the dynamic, adaptive parts of memory (sleep-dependent consolidation, emotional salience weighting, self-monitoring) are barely sketched. The gap is in knowing what to remember, when to forget, and how much to trust what you recall.

How the series is organized

Each post is a stand-alone read of about 900 words, structured the same way: the biology, the technology, then the gap. You can read them in order or jump straight to the component that is bothering you in production right now.

Short-term systems

How information enters the brain and what survives the first second:

• 1. Sensory Memory: The Quarter-Second Buffer Behind Whisper and Kafka
• 2. Working Memory and Context Windows: Why Your Agent Forgets Mid-Task
• 3. Attention Gating: From Broadbent's Filter to Sparse Transformers

Long-term stores

The three classic boxes from Tulving's classification, each with its own analog in production:

• 4. Episodic Memory: Time Cells, Place Cells, and Bi-Temporal Graphs
• 5. Semantic Memory: How Brains and Vector Databases Both Lose the Source
• 6. Procedural Memory: The Skills Layer AI Agents Are Just Starting to Build

The memory lifecycle

What happens to a memory between the moment it is formed and the moment it is recalled (or lost):

• 7. Memory Encoding: Why Deep Processing Beats Keyword Search
• 8. Memory Consolidation: How Sleep Inspired Background Memory Agents
• 9. Memory Retrieval: Pattern Completion in Hippocampus and Vector Search
• 10. Memory Reconsolidation: When Remembering Rewrites the Memory
• 11. Forgetting Is a Feature: Why Good Memory Systems Throw Things Away

The smart layers

Less obvious mechanisms that decide what gets remembered, how memories link to each other, and how confident the system is in its own recall:

• 12. Memory Scoring: How the Brain (and Mem0) Decide What Matters
• 13. Associative Memory: From Hopfield Nets to "Attention Is Hopfield"
• 14. Metamemory: Can Agents Know What They Don't Know?
• 15. Context-Dependent Retrieval: Why Memory Leaks Across Tenants
• 16. Schemas: How Prior Knowledge Quietly Distorts Every Memory

Special-purpose systems

Two pieces that do not fit cleanly into the lifecycle but matter a lot for real agents:

• 17. Prospective Memory: Teaching Agents to Remember Future Intentions
• 18. Hippocampal Indexing: The Most Direct Bio-to-Tech Parallel in AI Memory

Who this is for

If you are building an agent and you keep reaching for the same vector database trick to solve every memory problem, this series is for you. The brain figured out a long time ago that one mechanism cannot do all of memory's jobs. Sensory buffers, working memory, episodic stores, and procedural skills are different things for good reasons. Once you can see those reasons, the design choices in your own system get a lot clearer.

Start anywhere. The order is suggested, not required. And if you read one and want to argue, that is the point.