%20rotate(-45)'/%3e%3ccircle%20class='cls-2'%20cx='1605.42'%20cy='180'%20r='179.96'/%3e%3ccircle%20class='cls-2'%20cx='2051.52'%20cy='2840.24'%20r='179.96'%20transform='translate(-1407.4772%202282.5296)%20rotate(-45)'/%3e%3ccircle%20class='cls-2'%20cx='886.68'%20cy='2862.36'%20r='179.96'/%3e%3ccircle%20class='cls-2'%20cx='2848.41'%20cy='1467.96'%20r='179.96'/%3e%3ccircle%20class='cls-2'%20cx='180'%20cy='2182.96'%20r='179.96'/%3e%3ccircle%20class='cls-1'%20cx='1137.47'%20cy='1926.41'%20r='374.02'/%3e%3cline%20class='cls-1'%20x1='2848.41'%20y1='1467.96'%20x2='1511.59'%20y2='1826.16'/%3e%3cline%20class='cls-1'%20x1='763.36'%20y1='2026.65'%20x2='180'%20y2='2182.96'/%3e%3cline%20class='cls-1'%20x1='1401.99'%20y1='2190.91'%20x2='2051.52'%20y2='2840.24'/%3e%3cline%20class='cls-1'%20x1='872.88'%20y1='1661.97'%20x2='180'%20y2='969.3'/%3e%3cline%20class='cls-1'%20x1='886.68'%20y1='2862.36'%20x2='1040.67'%20y2='2287.68'/%3e%3cline%20class='cls-1'%20x1='1234.28'%20y1='1565.13'%20x2='1605.42'%20y2='180'/%3e%3c/svg%3e)
Notes on portable context, graph retrieval, and the memory layer for AI.
Long-form thinking from the team building Memplex, the Portable Context System for AI. Architecture, research, and the case for moving memory out of the model.
- 7 min
Building MCP Servers: Lessons from Designing Memplex
MCP servers are easy to ship and hard to ship well. Five design patterns we landed on while building Memplex's, with reasoning for each.
MCPProtocol DesignDeveloper Tools - 6 min
Memory as the New Moat: Vendor Lock-In in the AI Era
Model quality is converging. Memory isn't. The vendor that holds your context for six months has effectively locked you in — and that's the actual competitive dynamic of the next AI era.
AI StrategyCompetitionMarket Analysis - 6 min
Atlassian's Graph-RAG Numbers, Decoded
44% more accurate, 48% fewer tokens. Atlassian's published comparison of graph-RAG against vector RAG is the most important AI memory result of the past year. Here's what it actually measured and what it implies.
RAGResearchAI Architecture - 6 min
The Long-Horizon Agent Problem and the Persistent Context Layer
AI agents now run for hours, taking dozens of autonomous steps. Their memory hasn't caught up. The persistent context layer is what stands between agents that work and agents that lose track.
AI AgentsArchitectureMemory - 5 min
Why Vector RAG Is a Stopgap, Not a Future
Vector RAG is the default architecture for AI memory because it's the easiest to ship, not because it's the best. The fundamental limitation is that relationships matter and similarity doesn't capture them.
RAGAI ArchitectureRetrieval - 7 min
The Architecture Behind Memplex v0
A walkthrough of what's actually shipping in v0: code-signed Desktop Agent, Personal Graph, hybrid retrieval, MCP endpoint with explain_inclusion. And the things we're explicitly leaving out.
Build in PublicArchitectureMCP - 6 min
A Day in the Life: What Cross-Tool AI Memory Actually Looks Like
The abstract pitch for cross-vendor AI memory is straightforward. The concrete experience is more useful. Here's a walked-through day with Memplex in production, scoped routing and all.
Use CasesDeveloper ExperienceAI Memory - 5 min
The Personal vs Work AI Context Leak Problem
Your work AI shouldn't surface what you wrote on a Saturday night. Your personal AI shouldn't surface confidential project decisions. Current AI memory has no concept of scope. Here's what scope actually requires.
PrivacyAIEnterprise - 6 min
Why AI Memory Will Be Decoupled From AI Models
Storage decoupled from compute. Query decoupled from execution. AI memory is the next layer to be unbundled from the model it currently sits inside — and that's where a market gets created.
AI StrategyInfrastructureMarket Analysis - 5 min
User-Owned AI Context: The Real Privacy Story for the AI Era
You don't actually know what your AI tools currently remember about you. Each vendor has its own retention, its own export, its own access controls. The real fix isn't better disclosure — it's user-owned infrastructure.
PrivacyData OwnershipAI - 6 min
The Three Retrieval Channels Every Serious AI Memory System Needs
Vector-only retrieval is one tool solving every problem. Memplex uses three retrieval channels — lexical, semantic, graph — and weights them per query. Here's why a single channel always falls short.
RetrievalRAGAI Architecture - 6 min
Scoped Routing: How Memplex Decides What an AI Tool Gets to See
Routing is the most consequential part of an AI memory system, and the part fewest people think about. Here's how Memplex resolves nine variables before returning a single byte of context.
AI ArchitecturePrivacyMCP - 5 min
Source Lineage: The Foundation of Trustworthy AI Memory
If an AI memory can't be traced back to a source, it's just confident hallucination. Lineage isn't a feature — it's the property that turns a memory system from a guessing machine into a system of record.
AI TrustProvenanceEnterprise AI - 6 min
explain_inclusion: Making AI Retrieval Auditable by Default
Every AI memory system today is a black box: context goes in, retrieval comes out, and the user has no way to know why. We think auditable retrieval should be the default, not the premium feature.
AI TrustObservabilityMCP - 5 min
Why We Will Never Ship an Upload Widget
If your AI memory product needs you to upload files, it isn't doing its job. The right ingestion model is autonomous — and we're committed to it as a product principle, not a future feature.
Product DesignAI MemoryBuild in Public - 7 min
What 'Context' Actually Means in MCP
Most MCP servers expose a flat resource list and call it done. That's the wrong abstraction. Real context is scoped, lineage-bearing, and route-aware — three things the protocol gives you room to do, and most servers don't.
MCPProtocol DesignAI Memory - 6 min
Cross-Vendor AI Memory: The Opening No Vendor Can Address
Every AI vendor is racing to own your memory. None of them will ever ship great integration with their competitors. That structural gap is the opening for neutral infrastructure.
AI StrategyInfrastructureMCP - 5 min
The Hidden Cost of Cold-Start AI Sessions
Every AI session today starts from zero. Multiply that across tools and weeks and the cost stops being trivial — it's hours per person per week, paid in cognitive friction.
AIProductivityDeveloper Experience - 6 min
Why Graph Retrieval Beats Vector RAG for AI Memory
Vector RAG returns what's semantically near. A graph returns what's structurally connected. For AI memory, the second is almost always what you want — and the published numbers back it up.
RAGKnowledge GraphAI MemoryRetrieval