Your AI Assistant's Memory Isn't Actually Yours
Three questions show why your AI coding assistant's built-in memory isn't a durable record — portable, synced, versioned — plus a worked audit proving it.
Rajiv Pant
Rajiv Pant - Technology Leader, Writer, Thinker
Latest posts
Slop detection, not AI detection
A free, open-source tool that reads content along two axes: AI-provenance signal and slop-independence. Built for editors, writers, and anyone deciding what is worth publishing.
The architecture of an open-source slop detector
Multi-pass orchestrator, seven-layer hosted-tier safety, tools-router pattern, and the open-source license choices that keep a free LLM-mediated tool alive. The engineering behind slopcheck.
A compounding pattern catalog for AI slop
Why a methodology for detecting AI-generated content should keep the patterns it retires, not delete them. The shape of the catalog behind slopcheck.
Three Ragbot releases, one architectural shift
Reading Ragbot v3.3 (local Gemma 4 as first-class), v3.4 (the conversational-runtime release), and v3.5 (substrate cleanup) as one architectural arc — Ragbot becomes the conversational reference runtime of synthesis engineering.
The Direction Dynamic in ops: driving multi-tool recovery through a single AI session
The technical fix was the easy part. The recovery was tedious GUI work across Search Console, Bing Webmaster, and Cloudflare. I drove all of it through one Claude Code session using MCP browser automation. The Direction Dynamic (the synthesis pattern of human direction and AI execution) applies to ops work as cleanly as it applies to code.
How static-site migrations silently break SEO, and the build-time check that catches it
A silent sitemap regression after my Hugo-to-Astro migration degraded indexing for three months. The diagnosis was small. The audit that followed was not. Here is the comprehensive fix and the build-time check that prevents recurrence.
Introducing synthesis writing
Introducing synthesis writing — the discipline of keeping the creative act of writing wholly human while letting AI help with the scaffolding. What it is, what it rejects, and what it sets out to do.
Postel's Law for the human-AI interface
When one end of a software contract is an LLM, Postel's Law isn't a courtesy — it's load-bearing. The shape that lets you ship: document-as-contract.
Synthesis Project Management Now Works Across Claude Code and Codex
Synthesis project management now works across Claude Code and OpenAI Codex by keeping durable project memory in CONTEXT.md, REFERENCE.md, and sessions. Updated July 2026 with the agent attribution convention: session logs record which agent did what, verified how.
Through the Stage Door
Fitz and I went to Broadway to watch his second cousin Emmet Smith perform as Albus Potter in Harry Potter and the Cursed Child. What happened after the show was even better.
Demo mode as a pattern for software
Ship every software product with built-in demo mode. The pattern serves authors, users, sales teams, and compliance — and synthesis engineering makes it cheap enough to be standard practice.
Restoring 16 Broken Blog Images Using AI, the Wayback Machine, and a 93GB Photo Backup
How I used Claude Code, the Wayback Machine, and AI vision to restore 16 broken images from a decommissioned Smugmug photo hosting service across 10 blog posts.
Synthesis Console: open-source tooling for synthesis engineering
An open-source local tool that renders project management markdown and YAML as browsable, searchable pages. Three dependencies, no client-side framework, and a thinking framework that shaped every architecture decision.
Transferable Knowledge: The Fifth Pillar of Synthesis Coding
You understand your AI-generated code. Can your teammates? Transferable Knowledge is the fifth pillar of synthesis coding -- ensuring AI-assisted work is comprehensible by engineers who were not present when it was generated, and that it travels across people, AI tools, the sibling synthesis crafts, time, and organizations.
"All tests pass" is a warning, not a clearance
Why a green test suite after a schema change should trigger suspicion, not confidence. A verification methodology that catches what tests structurally cannot — missing migrations, mock gaps, and the cognitive blind spots that affect humans and AI agents alike.
Stop asking me: configuring Claude Code permissions for uninterrupted flow
A practitioner guide to eliminating Claude Code approval prompt fatigue. Why granular command patterns fail for compound commands, the Bash(*) plus deny list approach that works, and ready-to-use settings.json templates.
Managing AI agent skills at scale: a three-repo architecture
A three-repo architecture for managing public, private, and team AI agent skills with provenance tracking, drift detection, and synthesis merge — the first standard for private skill collections.
The Synthesis Thinking Framework: A Practitioner's Guide
Five thinking modes for technical practitioners who work with AI agents daily. A detailed methodology with real examples from managing sixty concurrent AI collaboration projects.
Your AI Is a Thinking Partner. You're Using It as a Search Engine.
Most leaders ask AI for answers. The ones gaining a real edge ask it to think with them. Here's the five-mode framework that makes that partnership work.
Synthesis Skills: install methodology directly into your AI workflow
22 open-source Agent Skills encoding synthesis coding and synthesis engineering methodology. One command installs code review, content quality, project management, and more into Claude Code, Cursor, or Codex CLI.
Who keeps the AI dividend?
AI productivity gains flow to employers while workers get more work for the same pay. Evidence from BCG, NBER, and labor economics shows why sharing matters.
Synthesis merging: what happens after the first PR
Synthesis merging goes beyond conflict resolution. It is the creative act of integrating contributions into a coherent whole, making the combined result better than the sum of its parts.
The Tiered Context Architecture: Managing AI Working Memory at Scale
When AI context files grow to 1,000+ lines, they degrade the collaboration they were built to support. The tiered context architecture solves this with three information tiers, budget enforcement, and lifecycle management — essentially garbage collection for AI context.
Fred Mann
Frederick G. Mann hired me when I was a college kid in 1995 and changed the trajectory of my life. He passed away on February 13, 2026.
When a code review runbook becomes an upgrade playbook
I built a code review runbook to find problems. Then I used it on codebases I owned, and the review categories became execution phases. A checklist for assessment turned into a playbook for action.
What reviewing real codebases taught me about code review
I published a 900-check code review runbook. Then I used it on real engagements. The checklist worked, but the real lessons were about the process — starting with reviewing the wrong branch.
When Someone Else Contributes to Your Synthesis-Coded Project
Standard merge workflows break down for synthesis-coded projects. The adopt-and-adapt pattern provides a disciplined integration methodology that respects contributor work while maintaining quality.
No, I'm not doing bodybuilding
Some of my friends from Delhi think my daily fitness training is bodybuilding. It isn't. Here's how I explain the difference.
The hidden lesson of Richard Socher's career: AI makes engineering discipline more valuable, not less
Richard Socher built the research that powers today's AI coding assistants and then described exactly how organizations should use them. His 'managers of AI' framework and synthesis engineering arrive at the same conclusion.
The research lineage from recursive neural networks to human-AI collaborative development
Tracing the intellectual thread from Richard Socher's compositional representations through DecaNLP to the systems-level engineering challenges of human-AI software development.
From learned representations to engineered context: what Richard Socher's research tells us about working with AI
Richard Socher's research arc — from GloVe to DecaNLP to 'managers of AI' — maps directly onto the principles behind synthesis coding. The connections are structural, not cosmetic.
Code Review That Scales: An Open-Source Runbook for Agentic Engineering
Most code review checklists are either too shallow or too heavy. This open-source runbook tiers review depth by project complexity — and is designed for agentic coding with tools like Claude Code.
Richard Socher nailed the future of work with AI
Richard Socher identified the skills that matter in the AI era: delegate clearly, specify requirements, build trust. These are learnable skills, not job titles.
The Part of Your Job AI Can't Do (And Why It Matters More Now)
There's a lot of anxiety about AI making engineering skills obsolete. But the most valuable part of what you do was never the typing. The judgment, the architecture, the knowing-what-to-build — that's becoming more important, not less.
When Two Frameworks Point the Same Direction
LangChain published Agent Engineering on December 9. I'd been developing Synthesis Engineering since early November. We arrived at remarkably similar conclusions. Here's what that convergence means.
RAG Architecture Lessons from Practice
Building a retrieval-augmented generation system revealed patterns the tutorials don't cover. Contractions break search. Not everything needs an LLM. Full document retrieval should be first-class.
Data Format Contracts for AI Pipelines
When tools interact, implicit assumptions break silently. Explicit data format contracts prevent the debugging sessions you don't have time for.
Context Loss and How to Prevent It
Why AI assistants forget what you told them, and practical strategies to prevent it. The difference between context that survives and context that gets lost.
From Tool to Team: Managing AI as Distributed Engineering
When you're running multiple AI sessions in parallel, you're not using a tool anymore — you're managing a team. The skills that matter shift accordingly.
The Foundation-First Pattern
Why the most effective AI-assisted projects start with humans writing code by hand. The counterintuitive pattern that makes everything else work.
Synthesis Project Management
A lightweight project management system for human-AI collaboration. Designed for context preservation across conversation sessions, context compaction events, multiple AI agents, and synced workstations. Updated July 2026 with the project naming convention and a sharper case against relying on tool-native memory; April 2026 for the phase-2 folder layout and cross-agent support for Claude Code and…
What is Synthesis Engineering?
Synthesis engineering is a professional discipline for human-AI collaboration on complex work. Not just better prompting — a systematic approach with princ...
How Claude's Memory Actually Works (And Why CLAUDE.md Matters)
Understanding Claude's context architecture changes how you work with it. The tradeoffs between automatic continuity and on-demand depth explain why persistent context files work.
Cracking a 25-Year-Old Password with Claude Code
While cleaning up old files, I found a password-protected Word document from 1999. I had no memory of the password. Claude Code and I cracked it together.
What OpenAI's Sora Build Teaches Us About Synthesis Coding
Four engineers shipped a #1 Play Store app in 28 days using AI agents. Their approach validates the patterns I've been writing about — and adds new ones worth studying.
Scaling Synthesis Coding in Your Organization
A practical guide for CTOs and engineering leaders: cost analysis, team rollout strategy, training curriculum, and metrics for adopting synthesis coding at s...
Building ownwords: A Synthesis Coding Case Study
What separates synthesis coding from vibe coding isn''t the tools — it''s the decisions. This technical case study walks through the architectural choices,...
A Stranger's Words
A theater reviewer wrote about my son's performance. She didn't know she was writing about my kid.
Why Synthesis Coding Still Writes Code in the Age of LLMs
The answer is not "just let the model do it" — and code matters more, not less, as AI gets dramatically better Last spring, my son Fitz and I had an aftern...