Technical writing on data engineering, AI workflows, and the tools we build along the way.
I have spent thirty-five years watching data tools get declared obsolete and replaced, six times over. The pattern that survives every cycle is the one worth betting on -- and it tells you exactly how to read the current pressure to rebuild everything for AI.
Read moreLangGraph is gaining real adoption for agentic AI workflows. But for most teams evaluating it, the question is not how to build a pipeline -- it is whether LangGraph is the right architecture for their problem, and what it actually takes to run in production.
Read moreAgentic AI is everywhere in vendor decks, but most teams cannot explain what it actually means for their data operations. This guide cuts through the hype with a practitioner's definition, a real architecture example, and a framework for deciding whether you need outside help.
Read moreEvery Claude session begins with amnesia. LoreConvo auto-save hooks fix that with zero friction for data engineers and AI practitioners.
Read moreLoreConvo & LoreDocs — local-first AI memory and knowledge vaults.
Part 2 of a series on building a Claude plugin -- lessons on marketplace review, dependency pinning, and the two-step PyPI release process for LoreConvo.
Read morePart 1 of a series on building a Claude plugin -- lessons on context management, session memory, and team knowledge sharing from developing LoreConvo.
Read moreTen scheduled AI agents, one shared memory layer: how LoreConvo and LoreDocs eliminated agent drift, and the practical patterns that made it work.
Read moreWhen you switch between AI coding tools, context gets locked in each one. Persistent local memory lets every tool read from the same vault of decisions.
Read moreMost AI memory tools default to vector embeddings. Here is why I chose SQLite and FTS5 instead, and what the trade-offs look like in a real agent fleet.
Read moreHermes Agent by NousResearch hit 153K GitHub stars in under three months by offering a compelling model-agnostic alternative to IDE-native AI tools. Here is how the two stacks compare when developer memory is the deciding factor.
Read moreEvery AI session starts from zero -- no prior decisions, no awareness of completed work. I built ten scheduled agents to measure the real cost.
Read moreImport your Obsidian vault into LoreDocs and query it from Claude, with full FTS5 + semantic search, versioning, and no cloud dependency.
Read moreGitHub Copilot, Claude Code, and Claude.ai all ship memory now. Memory is commodity. What is not commodity: the decisions you made last Tuesday, the open questions from your last design review, and the context that only makes sense if you know why -- not just what -- your agent did.
Read moreMem0 builds a knowledge graph by automatically extracting facts from every conversation. LoreConvo stores structured session narratives you explicitly commit. They are different bets about what 'memory' should do -- and the difference matters at $8/mo versus $249/mo.
Read moreA new wave of local-first AI memory tools has landed, and some of them require Docker, Postgres, and a vector database before you can save your first memory. Local data control and zero-ops simplicity are not the same thing. Here is what the difference looks like in practice.
Read moreLoreConvo launched on Product Hunt and Hacker News on May 27, 2026. It landed at position 208 with one upvote. Here is what the numbers actually showed, what the diagnosis is, and what we are changing before the next launch.
Read moreAnthropic's memory_20250818 tool gives your agent a filesystem and read/write operations. Everything else -- search, tagging, expiration, cross-surface access -- is left as an exercise for the developer. Here is what teams building on the primitive are reinventing, and how LoreConvo fits into that picture.
Read morePublished research from a major AI memory provider shows a 91% controlled benchmark collapsing to 49% effective accuracy at 30 days in production. The gap reveals a structural challenge with auto-capture memory systems -- and points to why explicit, structured saves may serve professional agents better than automatic accumulation.
Read moreThere are two ways to give an AI agent persistent memory: build a palace or let the agent take notes for you. The distinction matters more than which tool has the most stars.
Read moreEvery AI session captures how you think -- your business context, your half-formed ideas, your reasoning process. We built LoreConvo to keep that history on your machine, not in a SaaS vendor's database. Here is why the distinction matters.
Read moreWe pointed an OpenAI Codex agent at a project that uses LoreConvo. With no Codex-specific configuration, it registered LoreConvo as an MCP server, called save_session natively, and tagged the session correctly. The mechanism: one .mcp.json file that both Claude Code and Codex read automatically.
Read moreNotion's official Claude MCP integration puts 50 million Notion workspaces one step from your AI agent. LoreDocs takes the opposite approach: a knowledge vault built specifically for AI, not adapted for it. Here is why the direction matters.
Read moreClaude-Mem auto-captures everything and has tens of thousands of happy users. LoreConvo takes a different bet: intentional, project-based organization that scales with complex work. Here is how to tell which one fits your situation.
Read moreLoreConvo and LoreDocs -- persistent session memory and structured knowledge vaults for Claude -- are now live on the official Anthropic plugin marketplace. Here is what they do, how to install them, and why we built them local-first.
Read moreClaude now remembers things between sessions for free. But built-in memory stores fragments, not sessions. If you work across multiple tools, projects, or surfaces, you need something that travels with you.
Read moreOur full-text search was the right call to ship fast. Real data from 217 sessions reveals exactly where it needs help. Honest numbers on why we're moving from FTS5-only to hybrid search.
Read moreYour memory tool's benchmark score means nothing if it can not run in both Claude Code and Claude Cowork. We break down what matters.
Read moreI tried claude-mem, hit the edges of it on a real project, and built something that makes a different set of trade-offs. Here is an honest comparison.
Read moreThree powerful tools for knowledge work. Three different purposes. Here is how to use all three without duplication or confusion.
Read moreYour AI project knowledge scattered across multiple tools is a liability. LoreDocs organizes it in one place, versioned and searchable, so Claude and your team always work from current information.
Read moreEvery Claude session begins with amnesia. LoreConvo auto-save hooks fix that with zero friction for data engineers and AI practitioners.
Read moreData engineering, agentic workflows, and how we work.
A LangGraph graph can finish clean and still take the wrong path. State-transition observability tells you whether the result was right.
Read moreA decision framework for choosing LangGraph vs LangChain -- when the graph abstraction earns its complexity, anchored in a real 19-node finance pipeline.
Read moreHow to design AI systems that make authorization explicit, surface agent actions visibly, and give teams the audit trails they need to trust what's running in production.
Read moreWhen autonomous AI agents run on schedules, you need data to know what they actually did. We built a SQLite-backed observability pipeline for our multi-agent development team -- here is how it works and what we learned from the first weeks of real data.
Read moreI have spent thirty-five years watching data tools get declared obsolete and replaced, six times over. The pattern that survives every cycle is the one worth betting on -- and it tells you exactly how to read the current pressure to rebuild everything for AI.
Read moreWe built a three-agent AI workforce in early 2026. By spring, it had grown to ten agents running daily on schedules. Here is what broke when we scaled, the governance we had to build, and the honest lessons from six months of operating autonomous agents in production.
Read moreRoughly nine out of ten AI proofs of concept never become a service that runs on its own schedule. The gap between a tidy demo and a system that survives midnight alerts is a set of disciplined engineering decisions -- not a talent problem.
Read moreEveryone claims agentic AI expertise now. Here is how to tell the difference between someone who has shipped production systems and someone who has written demos. The questions to ask, the red flags to watch for, and an honest look at when you should just do it yourself.
Read moreA single mis-priced transaction can cost a firm millions; the maker-checker pattern gives you a reliable second opinion for AI-driven pipelines. We show how pairing deterministic code with an LLM auditor catches errors that each alone would miss.
Read moreWhen deterministic ETL pipelines start to crumble under ambiguous data and regulatory pressure, agentic workflows can restore reliability without a full rewrite.
Read moreLangGraph is gaining real adoption for agentic AI workflows. But for most teams evaluating it, the question is not how to build a pipeline -- it is whether LangGraph is the right architecture for their problem, and what it actually takes to run in production.
Read moreAgentic AI is everywhere in vendor decks, but most teams cannot explain what it actually means for their data operations. This guide cuts through the hype with a practitioner's definition, a real architecture example, and a framework for deciding whether you need outside help.
Read moreOne email when a new post is published — agentic AI, data engineering, and memory tools. No spam, no upsell, no AI summaries. Unsubscribe anytime.
Subscribe