https://modelbrain.net/articles/ Notes on AI memory, evaluation and local-first design from the ModelBrain team. en-us Wed, 28 Oct 2026 12:00:00 +0000 https://modelbrain.net/articles/chatgpt-claude-grok-export-formats.html https://modelbrain.net/articles/chatgpt-claude-grok-export-formats.html Wed, 28 Oct 2026 12:00:00 +0000 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. https://modelbrain.net/articles/export-chatgpt-conversations.html https://modelbrain.net/articles/export-chatgpt-conversations.html Wed, 28 Oct 2026 12:00:00 +0000 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. https://modelbrain.net/articles/how-to-delete-chatgpt-memory.html https://modelbrain.net/articles/how-to-delete-chatgpt-memory.html Mon, 26 Oct 2026 12:00:00 +0000 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. https://modelbrain.net/articles/chatgpt-cites-ai-memory-questions.html https://modelbrain.net/articles/chatgpt-cites-ai-memory-questions.html Mon, 26 Oct 2026 12:00:00 +0000 Seven AI-memory prompts sent to ChatGPT on October 26, 2026: every cited source was the vendor's own domain, none were independent publishers. https://modelbrain.net/articles/what-is-an-ai-second-brain.html https://modelbrain.net/articles/what-is-an-ai-second-brain.html Sat, 24 Oct 2026 14:15:00 +0000 An AI second brain stores your notes so an assistant can answer from them. Compare three setups by who does the filing. Checked Oct 24, 2026. https://modelbrain.net/articles/shared-ai-memory-small-team-someone-leaves.html https://modelbrain.net/articles/shared-ai-memory-small-team-someone-leaves.html Sat, 24 Oct 2026 14:10:00 +0000 Compare repository files, Claude team features and encrypted shared partitions on what a leaver keeps, what the team keeps and what they can no longer see. Checked Oct 24, 2026. https://modelbrain.net/articles/claude-memory-vs-chatgpt-memory.html https://modelbrain.net/articles/claude-memory-vs-chatgpt-memory.html Thu, 22 Oct 2026 14:45:00 +0000 Claude and ChatGPT both remember you. Compare them by what you can take out, in what format, and whether the other assistant can use it. Checked Oct 22, 2026. https://modelbrain.net/articles/can-chatgpt-use-a-memory-on-your-computer.html https://modelbrain.net/articles/can-chatgpt-use-a-memory-on-your-computer.html Thu, 22 Oct 2026 14:30:00 +0000 Yes, two ways: the ChatGPT desktop app runs local MCP servers, and ChatGPT on the web needs a tunnel or relay. Checked against OpenAI's docs on Oct 22, 2026. https://modelbrain.net/articles/encrypted-ai-memory-three-readers.html https://modelbrain.net/articles/encrypted-ai-memory-three-readers.html Mon, 19 Oct 2026 17:30:00 +0000 An AI memory can be read on the vendor's server, on your own disk and by the AI model. Here is how ModelBrain's three modes handle each reader, checked Oct 19, 2026. https://modelbrain.net/articles/claude-code-cursor-agents-md.html https://modelbrain.net/articles/claude-code-cursor-agents-md.html Fri, 16 Oct 2026 13:00:00 +0000 Yes, through AGENTS.md. Cursor reads it, and Claude Code reads it when no CLAUDE.md exists or when CLAUDE.md imports it. We ran three layouts on Claude Code 2.1.301. https://modelbrain.net/articles/claude-native-memory-vs-layer.html https://modelbrain.net/articles/claude-native-memory-vs-layer.html Fri, 16 Oct 2026 12:00:00 +0000 Claude has four memory stores. Compare them by who writes them, how far they reach and which other app can read them, checked against Anthropic and Cursor pages on Oct 16, 2026. https://modelbrain.net/articles/claude-read-obsidian-vault-obsidian-closed.html https://modelbrain.net/articles/claude-read-obsidian-vault-obsidian-closed.html Wed, 14 Oct 2026 13:00:00 +0000 Yes, if the tool reads the vault folder directly. Servers built as Obsidian plugins need Obsidian open. A one-minute closed-vault test, and which options pass. https://modelbrain.net/articles/obsidian-mcp-servers-compared.html https://modelbrain.net/articles/obsidian-mcp-servers-compared.html Wed, 14 Oct 2026 12:00:00 +0000 Obsidian MCP servers compared by write surface: how many tools can change or delete a note, whether a read-only switch exists, and if Obsidian must be open. https://modelbrain.net/articles/what-leaves-your-computer-ai-memory-tool.html https://modelbrain.net/articles/what-leaves-your-computer-ai-memory-tool.html Sun, 11 Oct 2026 13:00:00 +0000 A memory's text leaves your computer at two moments: when it is saved and when it is recalled. Recall always reaches your AI provider. Save time is where memory tools differ. https://modelbrain.net/articles/does-obsidian-have-ai.html https://modelbrain.net/articles/does-obsidian-have-ai.html Sun, 11 Oct 2026 12:00:00 +0000 Obsidian has no built-in AI assistant. Three ways to add one to your vault, what each one sends where, and how to connect Claude, ChatGPT and Cursor without a plugin. https://modelbrain.net/articles/persistent-memory-across-ai-assistants.html https://modelbrain.net/articles/persistent-memory-across-ai-assistants.html Wed, 07 Oct 2026 12:00:00 +0000 Each AI app keeps its own memory, so none of them know what you told the others. Here is how to give Claude, ChatGPT and Cursor one shared memory that stays on your computer. https://modelbrain.net/articles/what-accurate-should-mean.html https://modelbrain.net/articles/what-accurate-should-mean.html Sun, 04 Oct 2026 12:00:00 +0000 Why an AI memory tool's accuracy number means little without its scoring method: our two published numbers, 70% strict and 86% lenient, and the 79% weak spot. https://modelbrain.net/articles/sixseed-ceiling1200-benchmark.html https://modelbrain.net/articles/sixseed-ceiling1200-benchmark.html Sun, 04 Oct 2026 12:00:00 +0000 Six seeds, three arms, a full 285-row evaluation, and a pass/fail bar written down before we saw a single number. The longer ceiling cut truncation by two-thirds and still didn't ship. https://modelbrain.net/articles/vault-encryption.html https://modelbrain.net/articles/vault-encryption.html Mon, 14 Sep 2026 12:00:00 +0000 Local memory is protected by OS file permissions and kernel peer-credential checks, not by a cryptographic layer we invented. Here's why that's the right order. https://modelbrain.net/articles/per-node-pricing.html https://modelbrain.net/articles/per-node-pricing.html Mon, 14 Sep 2026 12:00:00 +0000 ModelBrain bills the machines that talk to each other, not the humans behind them, and a node's billable identity doesn't even exist until it needs to. https://modelbrain.net/articles/mcp-memory-servers-compared.html https://modelbrain.net/articles/mcp-memory-servers-compared.html Mon, 14 Sep 2026 12:00:00 +0000 Six MCP memory servers compared on storage, provenance tagging, and account requirements: the reference server, mcp-memory-service, Mem0, Letta, Zep, and ModelBrain. https://modelbrain.net/articles/local-first-rules-out.html https://modelbrain.net/articles/local-first-rules-out.html Mon, 14 Sep 2026 12:00:00 +0000 Local-first software is a specific, checkable set of guarantees, not a vibe. Here's exactly what ModelBrain's free tier rules out, and the one deliberate exception once you pay for device sync. https://modelbrain.net/articles/local-first-examples.html https://modelbrain.net/articles/local-first-examples.html Mon, 14 Sep 2026 12:00:00 +0000 Obsidian, Logseq, Actual Budget, Excalidraw and others, and what they actually guarantee. AI memory is almost always the exception to the pattern these tools follow. https://modelbrain.net/articles/how-we-tag-memory.html https://modelbrain.net/articles/how-we-tag-memory.html Mon, 14 Sep 2026 12:00:00 +0000 Every memory carries a provenance tier (source evidence, your own stated assertion, or a model-generated summary), kept distinct instead of collapsed into one opaque blob. https://modelbrain.net/articles/forget-guarantee.html https://modelbrain.net/articles/forget-guarantee.html Mon, 14 Sep 2026 12:00:00 +0000 A single forget() call that conflates undo, deletion, and permanent exclusion is a real bug class, not a hypothetical. Here's the explicit-mode guarantee that replaces it. https://modelbrain.net/articles/five-questions.html https://modelbrain.net/articles/five-questions.html Mon, 14 Sep 2026 12:00:00 +0000 A checklist for evaluating recall claims, including ours, before you rely on one with anything that matters. https://modelbrain.net/articles/five-calls-not-six.html https://modelbrain.net/articles/five-calls-not-six.html Mon, 14 Sep 2026 12:00:00 +0000 ModelBrain's MCP server exposed five tools in September 2026 (remember, recall, forget, expand, connector_sync) and why a wider tool surface isn't a missing feature. It now exposes fourteen; the current list is on the docs page. https://modelbrain.net/articles/designing-recall-test.html https://modelbrain.net/articles/designing-recall-test.html Mon, 14 Sep 2026 12:00:00 +0000 Before building more of the product, we wrote down how we'd know if it actually worked: five concrete, adversarial questions instead of a demo video. https://modelbrain.net/articles/backup-is-included.html https://modelbrain.net/articles/backup-is-included.html Mon, 14 Sep 2026 12:00:00 +0000 Managed backup is not a second product with a second subscription. It comes with device sync, and the only thing metered is the space it uses, at the same flat $0.05/GB/month as every other byte we store for you.