Articles
Notes on memory, evaluation and local-first design, written as we build rather than after. Short, and dated so you can tell what's changed since.
ChatGPT, Claude and Grok Export Formats Compared
Real exports from ChatGPT (1,463 conversations), Claude (223) and Grok (309) compared by structure and by how much of each file is conversation text.
How to Export ChatGPT Conversations and Read What Is Inside
Export ChatGPT data from Settings, then read it correctly: conversations are a tree, not a transcript. Findings from a real 1,463 conversation, 916 MB export.
How to Delete ChatGPT Memory, and Check That It Worked
Delete a saved memory, delete its chat, then run a two-query check. Includes a timed delete test: recalled, forgotten, 0 hits seven seconds later.
Which Sources ChatGPT Cites When You Ask About AI Memory
Seven AI-memory prompts sent to ChatGPT on October 26, 2026: every cited source was the vendor's own domain, none were independent publishers.
What is an AI second brain? Judge every setup by who does the filing
Three ways to build an AI second brain, compared on who captures a thought, who keeps it current and what the assistant can do with it.
Shared AI memory for a small team: what happens when someone leaves
Repository files, Claude team features and encrypted shared partitions compared on what a leaver keeps and what they can no longer see.
Claude memory vs ChatGPT memory: which one lets you take it with you
Compare ChatGPT memory and Claude memory by what you can take out, in what format, and whether the other assistant can use it.
Can ChatGPT use a memory that lives on your computer?
The ChatGPT desktop app, a tunnel and a relay compared on who runs the hop and whether ChatGPT can save or only read.
Encrypted AI memory: three readers to check before you trust one
The three readers a saved memory can have, which ModelBrain modes close each, and what stays open.
Claude memory explained: what each kind remembers and who can read it
Four memory stores, compared by who writes them, how far they reach and which other app can read them.
Can Claude Code and Cursor read the same instruction file?
A three-layout test on Claude Code 2.1.301: when AGENTS.md loads, when it is ignored, and how to share one file.
Obsidian MCP servers compared by what they can change
Count the tools that can change or delete a note, whether a read-only switch exists, and which servers need Obsidian open.
Can Claude read your Obsidian vault with Obsidian closed?
A one-minute test for whether an Obsidian tool needs the app open, and which options pass.
What leaves your computer when you use an AI memory tool
Text leaves at two moments, save and recall. Recall always reaches your AI; save time is where memory tools differ. Tested against ModelBrain and Mem0.
Does Obsidian have AI? Connect your vault to Claude, ChatGPT and Cursor
Obsidian has no built-in AI assistant. Three ways to add one to your vault, and what each one sends where.
How to give Claude, ChatGPT and Cursor the same persistent memory
Each AI app keeps its own memory. Here is the setup that gives them one shared memory on your own computer, and what it does not do yet.
What "accurate" should mean for an AI's memory
Why an accuracy number means little without its scoring method: our 70% strict and 86% lenient results, and the 79% weak spot, published together.
Why we didn't ship a 1200-character reasoning ceiling
Six seeds, three arms, a full 285-row evaluation, and a pass/fail bar written down before we saw a single number.
Most MCP memory servers store text. Few tag where it came from.
Six real MCP memory servers, compared on storage, provenance tagging, and whether an account is required.
Local-first software examples: where memory fits the pattern
Eight real local-first examples, Obsidian to Huly, and where an AI assistant's memory usually falls outside the pattern.
ModelBrain's MCP server tools: the full list of five calls
The September five-call surface (remember, recall, forget, expand, connector_sync) and why it was kept small. The server now exposes fourteen tools; see the docs.
How ModelBrain tags what it remembers
Source evidence, stated assertion, or model-generated summary, kept distinct instead of one opaque blob.
Why managed backup is included, not billed separately
It comes with device sync; the only thing metered is the space it uses, at the same $0.05/GB as everything else we store.
What Per-Node Pricing Means, and Why We Don't Count People
The free tier has no login for a seat count to attach to, so we bill machines, not people.
What a Forget Call Actually Has to Guarantee
Undo versus deletion, and why we test them as separate cases.
Designing the recall test before writing the product
Why the evaluation came before most of the features it measures.
What Local-First Software Actually Rules Out
Where the line is between local-first and merely local-by-default.
Why we didn't build vault encryption first
The tradeoff behind the decision, and what protects the local tier instead.
Five questions to ask any AI memory tool before you trust it with your work
A checklist for evaluating recall claims, including ours, before you rely on one.