Engineering Memory
Where cognitive science meets systems engineering.
Deep-dives into how memory actually works — in brains and in machines. Hebbian learning, forgetting curves, spreading activation, knowledge graphs, and the engineering required to make AI agents that genuinely remember.
Cognition & Neuroscience
How the brain remembers — and how we translated decades of cognitive science into code
What Is AI Agent Memory? Beyond Chat History and RAG
AI agents are everywhere — but most forget everything between runs. What is agent memory, how does it differ from RAG, and what does a real memory system look like?
What Is AI Memory? A Technical Guide for 2026
AI memory is not a database. It is not RAG. It is not context windows. This guide explains what AI memory actually is — and why every serious AI system will need one.
Cognitive Memory: The Missing Piece in the AI Arms Race
Every lab is racing on reasoning, planning, and tool use. But memory — the ability to learn from experience — is the capability nobody is shipping. That is about to change.
Giving OpenAI Agents Cognitive Memory: Shodh-Memory + Agents SDK
OpenAI's Agents SDK gives agents tools and handoffs. Shodh-memory gives them cognition — memory that strengthens with use, decays naturally, and surfaces context before you ask.
Building AI Agents That Actually Learn: Beyond Prompt Engineering
Most AI agents are stateless wrappers around LLMs. They don't learn — they restart. Here's what it takes to build agents with genuine cognitive continuity.
Vector Search Beyond Cosine Similarity: Graph-Based Approaches
Cosine similarity is chapter one. Graph-based search (Vamana, DiskANN) is chapter two. How shodh-memory auto-scales from HNSW to SPANN at 100K vectors.
Cognitive Architecture for AI Systems: What Neuroscience Actually Teaches Us
Baddeley's working memory. Cowan's embedded processes. Ebbinghaus forgetting curves. Hebb's rule. How each maps to engineering decisions in a real system.
Memory in Multi-Agent Systems: When AI Agents Share a Brain
When multiple agents share memory, you get emergent coordination. When they don't, you get chaos. The architecture of shared cognitive systems.
The Memory Layer: Why Every AI System Will Have One by 2027
Compute has a layer. Storage has a layer. Networking has a layer. Memory is the missing layer in AI infrastructure — and it's arriving now.
Hebbian Learning for AI Agents: Neurons That Fire Together Wire Together
How we implemented biological learning principles in shodh-memory. When memories are accessed together, their connection strengthens—just like synapses in the brain.
The Three-Tier Memory Architecture: From Cowan to Code
Deep dive into our sensory buffer, working memory, and long-term memory tiers. Based on Nelson Cowan's embedded-processes model.
Memory Decay and Forgetting Curves: The Math Behind Remembering
Ebbinghaus showed us forgetting is predictable. We implement hybrid exponential + power-law decay for realistic memory behavior.
Knowledge Graphs and Spreading Activation: How Context Surfaces
When you access one memory, related concepts activate too. We implement spreading activation for proactive context retrieval.
Long-Term Potentiation in Code: Making Memories Permanent
In the brain, repeated activation makes synapses permanent. We implement LTP so frequently-used knowledge resists decay.
Architecture & Engineering
System design decisions, storage engines, and the infrastructure of memory
Why Robotics Still Doesn't Have Memory (And How to Fix It)
Robots can see, grasp, and navigate. But they can't remember what they learned yesterday. The robotics memory gap is real — and solving it requires rethinking how memory works.
MCP: The Protocol That Will Define How AI Tools Communicate
Model Context Protocol is to AI tools what HTTP was to the web. A deep dive into the protocol, its design, and why memory is its killer app.
Why We Chose Rust for AI Infrastructure (And When You Shouldn't)
An honest take on Rust for AI systems. The wins: memory safety, zero-cost abstractions, cross-compilation to ARM. The costs: iteration speed, ecosystem gaps.
RocksDB for AI Workloads: Lessons from Building a Memory Engine
Why we chose RocksDB over SQLite and Postgres. Column families, prefix iterators, write-ahead logging, and the compaction strategies that actually matter.
Running Embedding Models on Edge Devices: ONNX, Quantization, and Reality
Getting MiniLM-L6-v2 to run on a Raspberry Pi at 34ms per embedding. ONNX Runtime, model quantization, batch processing, and the circuit breaker that saved us.
RAG Is Not Memory: Why Your AI Still Has Amnesia
Everyone thinks RAG solves the memory problem. It doesn't. Retrieval is not remembering. Here's the difference—and why it matters.
Memory Architecture for Autonomous Agents: Why Your AI Needs a Brain, Not a Database
Autonomous agents are everywhere—coding assistants, research bots, robotic systems. But most are goldfish. Here's how to give your agent a real brain.
Why Vector Search Alone Isn't Enough for Agent Memory
Vector similarity is great, but agents need more. We explain why shodh-memory combines vectors with knowledge graphs and temporal indices.
Benchmarking AI Memory Systems: Latency, Accuracy, and Scale
How does shodh-memory compare to alternatives? We share our benchmarking methodology and results across key metrics.
AI Agents & MCP
Building agents that remember, learn, and coordinate through shared cognitive systems
Why Your Coding Assistant Forgets Everything: Fixing AI Memory for Developers
You've explained your project structure 47 times this month. Your AI assistant has the memory of a goldfish. Here's how to fix that.
The Agentic Shift: Why 2026 Is the Year AI Stops Waiting for Prompts
We're witnessing the biggest shift in AI since ChatGPT. Agents that act, remember, and learn—not chatbots that wait. Here's what's actually changing.
Integrating shodh-memory with Claude Code and Cursor via MCP
Complete guide to adding persistent memory to your AI coding assistant. One command to remember everything across sessions.
Edge & Robotics
Running cognitive memory on Raspberry Pis, robots, and air-gapped systems
Why AI Memory Should Run Locally: Privacy, Latency & Sovereignty
Cloud AI memory means your agent's knowledge lives on someone else's server. Here's why local-first memory wins on privacy, speed, and control — and how to do it.
Running AI Memory on a Raspberry Pi: A Practical Guide
Step-by-step guide to deploying shodh-memory on resource-constrained devices. Achieve sub-100ms semantic search on a $35 computer.
Memory for Robots: Learning from the Real World
Case study: how a pick-and-place robot uses shodh-memory to learn object positions and adapt to warehouse changes.
Privacy-First AI Memory: Why Your Data Should Stay Local
In an age of cloud AI, we argue for local-first memory. Your agent's knowledge is valuable—keep it on your hardware.
31 posts on cognition, memory architecture, and AI systems