# ModelBrain™

> The private AI memory layer: one memory for every assistant you use (Claude, ChatGPT and Cursor), on every computer you use. It runs on your own computer, needs no account for local use, and what we store for you we cannot read. Published recall test: 269 questions run three times, 86% right by a lenient automatic match and 70% by exact match. Beta: pages here say plainly what's shipped versus only designed.

ModelBrain is not a chat interface or an assistant itself. It's an MCP (Model Context Protocol) server any compatible AI assistant can call to remember, recall, forget, and expand on things across sessions, with the memory stored as ordinary files on the user's own machine rather than in a vendor account.

Full text of the main pages in one file: [llms-full.txt](https://modelbrain.net/llms-full.txt), regenerated each release.

## Core pages

- [How it works](https://modelbrain.net/how-it-works.html): Plain explanation of the local daemon + MCP tool surface, then the architecture underneath.
- [Architecture](https://modelbrain.net/architecture.html): The four layers (assistant, the fourteen MCP tools, the local program, and where memories physically sit), with v1-shipped marked separately from designed-but-not-built.
- [Compare](https://modelbrain.net/compare.html): Comparison against Mem0 and Letta/MemGPT on where data lives by default, what forgetting does, and what recall-quality evidence exists.
- [Mem0 alternative for one person: ModelBrain vs Mem0](https://modelbrain.net/compare/mem0.html): A first-memory test of Mem0 open source (mem0ai 2.2.1, run Oct 19, 2026) against ModelBrain local: Mem0's default setup needs a runtime, an OpenAI key, an extraction model, an embedding model, a storage path and an identifier before it saves one memory, and failed at setup without a key. ModelBrain local needs no account or key. Also says where Mem0 is the better choice.
- [Pricing](https://modelbrain.net/pricing.html): Free for local use, permanently. Paid tiers are per-node device sync at $6 per device a month billed yearly or $8 monthly, $4 a month per device past 25 on either plan, with managed backup included and stored data at $0.05/GB a month, available now on the paid plan; Link, a Stripe service, is the merchant of record and handles sales tax, VAT and GST; Teams (available now: shared partitions, end-to-end encrypted, with roles, SSO and an audit log; start at https://modelbrain.net/account/org/) uses the same per-node rate on members' computers, and a computer already on its owner's sync plan isn't charged again. Connecting ChatGPT through the relay is free with a free account. Cancelling keeps access to the end of the paid period. Start syncing at https://modelbrain.net/account/ (choose yearly or monthly).
- [FAQ](https://modelbrain.net/faq.html): Account requirements, storage location, what "forgotten" actually guarantees, cost, and ship timing.
- [Security](https://modelbrain.net/security.html): Trust boundaries, what is and isn't encrypted, what ChatGPT and Grok can read through the relay (private imported sources stay hidden unless you allow a partition), what the Claude Code hook sends to Anthropic, and what signing in sends. Updated Oct 28, 2026 for 0.6.6.
- [Privacy](https://modelbrain.net/privacy.html): What's held in each mode (local, sync and backup, Teams), what stays encrypted so we can't read it, and what's never used as training data.
- [Docs](https://modelbrain.net/docs.html): The fourteen MCP tools (remember, recall, list_memories, expand, forget, six partition tools, diagnostics, usage, ping), the complete surface.
- [Changelog](https://modelbrain.net/changelog.html): What shipped, what was cut, what's still open, in order. Latest: 0.9.0 on macOS, Windows and Linux, the release candidate for 1.0 (the Claude Code hook answers in about 10 ms, saves no longer wait behind a search, a full disk returns storage_full and the store reopens on its own, the same benchmark retrieval as 0.6.8); 0.6.8 before it (a bundled learned ranker orders recall results, source_date on every hit, the relative-date slot on by default); 0.6.6 before it (opt in sources for Claude Code, ChatGPT, Grok and Claude exports, mail, calendar, contacts, bookmarks and git history, with the reranker bundled and on by default); 0.6.5 before it (retries that never mark a file as forgotten, tighter checks on unidentified programs and relay grants, folder includes that keep the secret-file skip, and a tray action to resume a forgotten synced file); 0.6.3 before it on macOS, Windows and Linux (tighter checks on which programs may act as an assistant or ChatGPT relay, a per-minute and per-call limit on forgets, and a lighter store because forgotten folder content leaves only a small stub); 0.6.2 before it on macOS, Windows and Linux (fewer false positives in the secret and prompt-injection scans, per-client forget limits that survive reconnects, token revoke and a token label on each connected client); 0.6.1 before it on macOS, Windows and Linux (scale and safety release: shared reads, staged compaction, per-project partitions, undo for forgets, an Allow or Block prompt for new assistants; stores written by 0.6 do not open in 0.5.x); 0.5.1 added export and recall from the command line; 0.5.0 before it (setup for five more assistants, an optional as_of date on recall, a small ranking bonus for memories tagged role:user); 0.4.4 before it on macOS and Windows (a setup guide window in the tray app, no black console windows on Windows, ChatGPT stays connected, clearer ChatGPT sign-in error), and the previous Linux build; 0.4.3 before that (safer crash recovery, recall folds identical passages by default, synced folder text is purged 24 hours after it is forgotten, many hardening fixes).
- [Recall test results](https://modelbrain.net/results.html): The full published results: 269 questions run three times (807 scored answers), 86% right by a lenient automatic match and 70% strict, per-case rates (100% single fact, prompt injection and declining; 94% corrected facts; 90% buried and lookalike; 79% connecting two notes), the six-seed 285-question run, what is not counted, and what the test does not prove.
- [Benchmarks](https://modelbrain.net/benchmarks.html): ModelBrain 0.6.8, as its installers run it (cross-encoder reranker plus a bundled learned ranker), against Mem0 2.2.1, Honcho 2.7.0 and total-agent-memory (pinned commit 16d8c82, reranker off and on) on the held-out test splits of LoCoMo (1,152 questions), BEAM 100K (126), BEAM 500K (252) and LongMemEval (370), same scorer and per-question rows for every system. recall_any@5 / recall_all@5 for ModelBrain 0.6.8: 0.810/0.666 LoCoMo, 0.746/0.389 BEAM 100K, 0.603/0.314 BEAM 500K, 0.986/0.919 LongMemEval; Mem0 0.670/0.530, 0.611/0.262, 0.520/0.218, 0.973/0.873; Honcho 0.611/0.476, 0.651/0.325, 0.540/0.226, 0.951/0.832. Of 32 paired comparisons (four benchmarks, two measures, four competitor settings) ModelBrain is ahead in 31 and tied in 1 (LongMemEval any@5 against Mem0) at a 95 percent interval. LongMemEval answer accuracy with the gpt-4o reader and official judge: ModelBrain 0.6.8 320 of 370, Mem0 312, Honcho 302, total-agent-memory 299 (ahead of Honcho and total-agent-memory on a 95 percent interval, ahead of Mem0 on the number). Thesis: answer accuracy follows how often the reader sees every piece of evidence (recall_all@5), not just one piece. Caveats on the page: the test splits were used in earlier experiments, competitors ran retrieval only and untuned, ModelBrain reads the user-turn tag, and 0.6.8 ranks more slowly per query than 0.6.6 (median 0.29 to 0.80 s). Scores: https://modelbrain.net/benchmarks/068-scores.txt. The earlier 0.6.6 locked run stays on the page below the 0.6.8 results.
- [Fact sheet](https://modelbrain.net/facts.html): Platforms, current version, price, hardware/model tiers, and exactly what leaves your computer, in tables.

## Guides

- [Install ModelBrain](https://modelbrain.net/docs/install/): Pick macOS, Windows or Linux, with the installer for each, how each one connects to your assistants, and where the local vault lives on each system.
- [Install on macOS](https://modelbrain.net/docs/install/macos.html): Download the .pkg, run the installer, and where the local vault lives on disk.
- [Install on Windows](https://modelbrain.net/docs/install/windows.html): Download the installer (signed by NexSpark LLC), install, and where the local vault lives; what to do if SmartScreen still warns right after a release.
- [Install on Linux](https://modelbrain.net/docs/install/linux.html): Download the tarball, run install.sh, command line only, x86_64 and aarch64, glibc 2.28 or newer, and where the local vault lives.
- [Connect Claude Desktop](https://modelbrain.net/docs/connect/claude-desktop.html): Self-registers automatically on launch; no manual JSON editing.
- [Connect Claude Code](https://modelbrain.net/docs/connect/claude-code.html): One button in ModelBrain's in-app Setup guide window.
- [Connect Cursor](https://modelbrain.net/docs/connect/cursor.html): Same one-button Setup guide flow, for Cursor.
- [Connect Codex CLI](https://modelbrain.net/docs/connect/codex-cli.html): Manual setup: add one [mcp_servers.modelbrain] table to config.toml. Not yet hand-tested on a real install.
- [Connect Gemini CLI (existing installs)](https://modelbrain.net/docs/connect/gemini-cli.html): Google replaced Gemini CLI with Antigravity CLI on 18 June 2026; this guide is kept for existing installs. Manual setup: add one mcpServers entry to settings.json, with trust left off, then check with gemini mcp list. Connection hand-checked on Gemini CLI 0.64.0 on macOS; remember and recall not yet tested, and Google sign-in is unavailable to individuals.
- [Connect Cline](https://modelbrain.net/docs/connect/cline.html): Manual setup: open Cline's MCP settings from inside Cline and add one entry. Not yet hand-tested on a real install.
- [Connect Antigravity](https://modelbrain.net/docs/connect/antigravity.html): Manual setup: add one mcpServers entry to ~/.gemini/config/mcp_config.json, which starts out empty. Hand-tested on a Mac: ModelBrain loaded and worked.
- [Connect OpenCode](https://modelbrain.net/docs/connect/opencode.html): Manual setup: add one entry under the mcp key in opencode.json, with the command as a list. Not yet hand-tested.
- [Connect Continue](https://modelbrain.net/docs/connect/continue.html): Manual setup: add one YAML file in .continue/mcpServers and use agent mode, the only mode that uses MCP. Not yet hand-tested.
- [Connect Hermes](https://modelbrain.net/docs/connect/hermes.html): Manual setup: add one entry under mcp_servers in ~/.hermes/config.yaml, then /reload-mcp. Not yet hand-tested.
- [Connect Devin Desktop](https://modelbrain.net/docs/connect/devin-desktop.html): Manual setup: add one entry to ~/.codeium/windsurf/mcp_config.json, the Windsurf-named folder Devin Desktop still reads, then restart. Cascade agent only. Not yet hand-tested.
- [Aider and ModelBrain](https://modelbrain.net/docs/connect/aider.html): Aider has no MCP command, so ModelBrain cannot connect to it. What to do instead.
- [Connect VS Code](https://modelbrain.net/docs/connect/vscode.html): Manual setup: add one entry under servers (not mcpServers) in VS Code's mcp.json, opened with MCP: Open User Configuration. Not yet hand-tested.
- [Connect Zed](https://modelbrain.net/docs/connect/zed.html): Manual setup: add one entry under context_servers in Zed's settings file; Zed asks before tools run by default. Not yet hand-tested.
- [Connect Goose](https://modelbrain.net/docs/connect/goose.html): Setup: run goose configure and add a command-line extension, or add one entry under extensions in ~/.config/goose/config.yaml. Not yet hand-tested.
- [Connect Roo Code](https://modelbrain.net/docs/connect/roo-code.html): Manual setup: open the global MCP file from Roo Code's MCP panel and add one entry, leaving alwaysAllow empty. Not yet hand-tested.
- [Connect LM Studio](https://modelbrain.net/docs/connect/lm-studio.html): Manual setup: add one entry to ~/.lmstudio/mcp.json, opened from the Program tab; use a model that can call tools. Not yet hand-tested.
- [ModelBrain and ChatGPT](https://modelbrain.net/docs/connect/chatgpt.html): Step-by-step ChatGPT setup: choose Connect to ChatGPT in the ModelBrain menu, turn on Developer mode in ChatGPT, add ModelBrain as a plugin (Settings, Plugins) with the relay URL and OAuth, then sign in to your ModelBrain account on modelbrain.net and pick the computer (no pairing code, nothing to allow on the computer). Each computer you connect needs a different name. If ChatGPT reports an old registration, delete the plugin and add it again. Needs a free ModelBrain account (the relay is free); disconnect from the Account window or the account page. Also covers how this one connection differs from the direct local connections Claude Desktop, Claude Code and Cursor use.
- [ModelBrain and Grok](https://modelbrain.net/docs/connect/grok.html): Step-by-step Grok setup (new): choose Connect to ChatGPT in the ModelBrain menu to turn on the relay for the computer, add a custom connector at grok.com/connectors with https://relay.modelbrain.net/mcp, then sign in to your ModelBrain account and pick the computer. Same relay as ChatGPT; Grok screen wording may differ.
- [Index Claude Code memory in ModelBrain](https://modelbrain.net/docs/connect/claude-code-memory.html): Add Claude Code memory folders and CLAUDE.md files to ModelBrain as a source. Available from ModelBrain 0.6.6: what is read, what is skipped, limits and how to check it.
- [Index your ChatGPT export in ModelBrain](https://modelbrain.net/docs/connect/chatgpt-export.html): Turn a ChatGPT data export into searchable conversations in ModelBrain. Available from ModelBrain 0.6.6: which messages are kept, do-not-remember handling, limits and privacy.
- [Index your Grok export in ModelBrain](https://modelbrain.net/docs/connect/grok-export.html): Turn a Grok data export into searchable conversations in ModelBrain. Available from ModelBrain 0.6.6: which messages are kept, temporary chats, limits and privacy.
- [Index your Claude export in ModelBrain](https://modelbrain.net/docs/connect/claude-export.html): Turn a Claude data export into searchable conversations in ModelBrain. Available from ModelBrain 0.6.6: which messages are kept, how large exports are streamed, limits and privacy.
- [Index mail from mbox files in ModelBrain](https://modelbrain.net/docs/connect/mail.html): Add mbox mail, Apple Mail bundles and Thunderbird folders to ModelBrain as a private source. Available from ModelBrain 0.6.6: what is kept, what is skipped, limits and privacy.
- [Index calendars and contacts in ModelBrain](https://modelbrain.net/docs/connect/calendar-contacts.html): Add .ics calendar files and .vcf contact files to ModelBrain as private sources. Available from ModelBrain 0.6.6: what is kept, what is skipped, limits and privacy.
- [Index browser bookmarks in ModelBrain](https://modelbrain.net/docs/connect/bookmarks.html): Add Chrome, Edge, Firefox and Safari bookmarks to ModelBrain as a source. Available from ModelBrain 0.6.6: one memory per bookmark folder, what is read, limits and privacy.
- [Index Git commit history in ModelBrain](https://modelbrain.net/docs/connect/git-history.html): Add a Git repository's commit messages to ModelBrain so you can ask why code changed. Available from ModelBrain 0.6.6: what each commit holds, the 5,000 commit default, limits and privacy.
- [The ModelBrain menu](https://modelbrain.net/docs/tray-menu.html): Reference for every item in the menu bar (macOS) and system tray (Windows) menu: status line, Copy Diagnostics, Setup guide, vault controls (Pause, Stop Syncing, Remove and Forget), Usage & Stats, the Account window, the Connect to ChatGPT states (Ready, Connected, Stop Sharing with ChatGPT, and the sign-in and plan states), reasoning toggles, Restart ModelBrain, and Quit Menu Bar Helper; from 0.6 also Undo recent forgets (15 minute window) and the Allow or Block window for a newly connecting assistant.
- [Clients and permissions](https://modelbrain.net/docs/clients-and-permissions.html): The Allow or Block prompt for new assistants, the forget limit of 20 objects a minute and per call, undo for 15 minutes, client tokens (mint, revoke, Require client tokens), which admin actions are refused, and the limit that client identity is a consent tool and not a defence against software already running as you.
- [Your ModelBrain account](https://modelbrain.net/docs/account.html): How to sign in (an emailed code, or Google, Microsoft or Apple on the account page) and what an account adds (sync between your computers with end-to-end encryption and managed backup and restore on the Sync plan; connecting ChatGPT, free), adding a computer with six matching words, billing, cancelling (access until the end of the paid period, then 30 days read-only before deletion), deleting the account, and the activity history.
- [Teams](https://modelbrain.net/docs/teams.html): Sharing a partition with a team: starting an organization, inviting and approving members (six team words compared in the app), roles (owner, admin, member, reader), per-partition read or write access, end-to-end encryption with per-partition keys only members' computers hold, removing someone (new key on the next sync; removals made on the web are confirmed in an owner or admin app before they are signed into the org history), single sign-on with OpenID Connect and a verified domain, the audit log, and billing per node.
- [Quickstart](https://modelbrain.net/docs/quickstart.html): Install, connect a client, add a folder, and verify recall in five steps.
- [Instruction tips](https://modelbrain.net/docs/instructions.html): Paste-ready project instructions and ten tips for getting an assistant to recall, save, label, partition and forget with ModelBrain.
- [Add, remove and rescan folders](https://modelbrain.net/docs/folders.html): What ModelBrain reads, what it automatically excludes (from 0.6 that includes credential-looking lines inside ordinary files, replaced with a redaction marker), and how to pause or remove a folder.
- [Export and search from the command line](https://modelbrain.net/docs/cli.html): modelbrain_admin export writes every saved memory to Markdown files (one per memory, a folder per partition) or one memories.jsonl, never overwrites, writes only the current version of each memory unless you add --history (0.6: replaced versions with superseded_by, and deleted_at for forgotten ones), skips synced folder chunks unless asked, and needs the app running; modelbrain_admin recall searches from a terminal; where the tool lives on macOS, Windows and Linux, exit codes, and what was and was not checked. Needs 0.5.1 or later.
- [Use ModelBrain with an Obsidian vault](https://modelbrain.net/docs/obsidian.html): Add an Obsidian vault folder (Add Vault in the menu, or connector_sync sync), what is read and skipped (the .obsidian folder and other dot-folders are skipped), background first pass, edits replace old text, deletions drop on a later sync, keeping private notes out, answers cite the source note, wikilinks, frontmatter, tags, embeds, canvas files and aliases are handled, a plain list of what is tested and what is not, no plugin and nothing leaves the computer for local use. Not affiliated with Obsidian.
- [Hardware requirements & model tiers](https://modelbrain.net/docs/hardware.html): RAM/CPU needed per reasoning tier, and how the model is chosen automatically.
- [Troubleshooting](https://modelbrain.net/docs/troubleshooting.html): Fixes for the most common install and connection issues, per platform.
- [Reasoning mode, explained](https://modelbrain.net/docs/reasoning.html): How the optional reasoning pass on recall differs from plain ranked search, what it costs, and the full list of reasons it can decline to answer rather than guess.

## Articles

- [Articles index](https://modelbrain.net/articles/): Full list, newest first.
- [Articles RSS feed](https://modelbrain.net/feed.xml): All articles with titles, summaries and dates, newest first.
- [Obsidian MCP servers compared by what they can change](https://modelbrain.net/articles/obsidian-mcp-servers-compared.html): Compares Obsidian MCP servers by write surface, the number of tools that can create, change or delete a note, plus read-only switches and whether Obsidian must be open, from each project's own page on Oct 14, 2026. ModelBrain's write surface on note files is zero.
- [Can Claude read your Obsidian vault with Obsidian closed?](https://modelbrain.net/articles/claude-read-obsidian-vault-obsidian-closed.html): Yes, when the tool reads the vault folder directly; plugin-based servers need Obsidian open. Gives a one-minute closed-vault test and a table of which options pass, with a limited first-party check on Oct 14, 2026.
- [What leaves your computer when you use an AI memory tool](https://modelbrain.net/articles/what-leaves-your-computer-ai-memory-tool.html): A memory's text leaves your computer at two moments, save and recall; recall always reaches the AI provider, so the save-time test is what separates memory tools. Covers ModelBrain with Claude, Cursor and ChatGPT, sync, and Mem0 defaults.
- [Does Obsidian have AI? Connect your vault to Claude, ChatGPT and Cursor](https://modelbrain.net/articles/does-obsidian-have-ai.html): Obsidian has no built-in AI assistant; the three ways to add one (plugin, MCP plugin, folder reader), what each sends where, and how to connect a vault to Claude, ChatGPT and Cursor, with the measured search speed and the limits.
- [How to give Claude, ChatGPT and Cursor the same persistent memory](https://modelbrain.net/articles/persistent-memory-across-ai-assistants.html): Why each AI app forgets, and the setup that gives Claude Desktop, Claude Code, Cursor and ChatGPT one shared memory on your own computer, plus how to make assistants use it, the published recall numbers, and what it does not do yet.
- [What "accurate" should mean for an AI's memory](https://modelbrain.net/articles/what-accurate-should-mean.html): Why an accuracy number means little without its scoring method: the 70% strict and 86% lenient results, and the 79% weakest category published alongside them.
- [Why we didn't ship a 1200-character reasoning ceiling](https://modelbrain.net/articles/sixseed-ceiling1200-benchmark.html): Six seeds, three arms, a full 285-row evaluation against the real production prompt path, and a pass/fail bar (including a hard latency bound) written down before the run.
- [Can Claude Code and Cursor read the same instruction file?](https://modelbrain.net/articles/claude-code-cursor-agents-md.html): Yes, through AGENTS.md. Cursor reads it; Claude Code reads it when no CLAUDE.md exists or when CLAUDE.md imports it. Includes a three-layout test run on Claude Code 2.1.301 on Oct 16, 2026.
- [Encrypted AI memory: three readers to check before you trust one](https://modelbrain.net/articles/encrypted-ai-memory-three-readers.html): A memory can be read on the vendor's server, on your own disk and by the AI model. Scores ModelBrain's local, sync and backup, and Teams modes on each reader: the server is closed for sync and Teams, the local disk is not encrypted at rest, and the model reader is open in every mode. Includes three questions to ask any memory tool.
- [ChatGPT Temporary Chat: What It Turns Off and What It Leaves On](https://modelbrain.net/articles/chatgpt-temporary-chat-what-stays-on.html): How to start a ChatGPT temporary chat, what it keeps out of history and training, and the three things it leaves on: your memory choice, a 30 day safety copy and third party actions.
- [How to Delete ChatGPT History, and What a Delete Leaves Behind](https://modelbrain.net/articles/how-to-delete-chatgpt-history.html): Delete one chat or clear all ChatGPT history, on web, iPhone and Android, and see which copies a delete does not reach, from files to exports you already downloaded.
- [ChatGPT Data Export Not Working: Check the Manifest Before You Retry](https://modelbrain.net/articles/chatgpt-data-export-not-working.html): Fix a ChatGPT export that is missing, expired or looks empty. A real 916 MB export had no conversations.json, and export_manifest.json is the receipt to check.
- [How Big Is Your AI Chat History? 104 Chunks per Conversation in ChatGPT, 375 in Claude](https://modelbrain.net/articles/ai-history-size-chunks-per-conversation.html): Real exports indexed on ModelBrain 0.6.6: ChatGPT 104 chunks per conversation, Claude 375, Grok 31. Why conversation count is a poor measure of size.
- [How to Move ChatGPT History to Claude, Memory and Conversations](https://modelbrain.net/articles/move-chatgpt-history-to-claude.html): Move ChatGPT memory with Claude import and conversations with the export zip. A real 1,462 conversation export indexed into 152,760 chunks in 118 seconds.
- [Search Your ChatGPT History: 82 Percent of It Is Not What You Wrote](https://modelbrain.net/articles/chatgpt-history-search-assistant-text-outweighs-yours.html): In a real ChatGPT export of 1,462 conversations, 82.1 percent of the 152,760 chunks were written by the assistant. Counts for Claude and Grok exports too.
- [How to Export Claude Conversations, and What the Seven Files Hold](https://modelbrain.net/articles/how-to-export-claude-conversations.html): Export Claude data from Settings, Privacy. A real export came as seven separate zip parts, with memory in its own part. Counts from a 272 conversation file.
- [How to Export Grok Conversations, and What Is Inside the File](https://modelbrain.net/articles/how-to-export-grok-conversations.html): Export Grok data from Settings, Data Controls. In a real 309 conversation export the assistant is written three ways, so match the sender case-insensitively.
- [How to Export ChatGPT Conversations and Read What Is Inside](https://modelbrain.net/articles/export-chatgpt-conversations.html): 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.
- [ChatGPT, Claude and Grok Export Formats Compared](https://modelbrain.net/articles/chatgpt-claude-grok-export-formats.html): Real exports from ChatGPT (1,463 conversations), Claude (272) and Grok (309) compared by structure and by how much of each file is conversation text.
- [How to Delete ChatGPT Memory, and Check That It Worked](https://modelbrain.net/articles/how-to-delete-chatgpt-memory.html): 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](https://modelbrain.net/articles/chatgpt-cites-ai-memory-questions.html): 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](https://modelbrain.net/articles/what-is-an-ai-second-brain.html): Defines an AI second brain and scores three setups on the filing test: who captures a thought, who keeps it current, and what the assistant can do with it. Vendor descriptions from MindStudio (May 27, 2026) and Tana (July 14, 2026), summarized and untested. ModelBrain reads a folder of notes you approve and stores what your assistant remembers; it does not yet detect contradictions between notes.
- [Shared AI memory for a small team: what happens when someone leaves](https://modelbrain.net/articles/shared-ai-memory-small-team-someone-leaves.html): The leaver test for team AI memory: what the leaver keeps, what the team keeps, and what they can no longer see. Compares repository files, Cursor rules, Claude Team or Enterprise removal (per Anthropic, memory behavior not stated) and ModelBrain Teams, where removal ends server access at once and the partition gets a new key at the next sync by a writer. Checked Oct 24, 2026.
- [Claude memory vs ChatGPT memory: which one lets you take it with you](https://modelbrain.net/articles/claude-memory-vs-chatgpt-memory.html): Compares ChatGPT memory, Claude memory and a memory you own on three exit questions: what you can take out, in what format, and whether another assistant can use it without retyping. Neither vendor offers a file export of memory (checked Oct 22, 2026); Claude imports pasted plain text, which is a summary of a summary. ModelBrain exports Markdown or JSON Lines since 0.5.1.
- [Can ChatGPT use a memory that lives on your computer?](https://modelbrain.net/articles/can-chatgpt-use-a-memory-on-your-computer.html): Yes, two ways. The ChatGPT desktop app runs local MCP servers directly; ChatGPT on the web cannot reach localhost and needs OpenAI's Secure MCP Tunnel or a hosted relay. Scores the routes on who runs the hop and whether ChatGPT can write or only read, with Pro limited to read and fetch for custom MCP per OpenAI. Checked Oct 22, 2026.
- [Claude memory explained: what each kind remembers and who can read it](https://modelbrain.net/articles/claude-native-memory-vs-layer.html): Claude has four memory stores: chat memory, CLAUDE.md, Claude Code auto memory and an outside memory layer. Compares them by who writes each, how far it reaches and which other app can read it, from Anthropic and Cursor pages checked Oct 16, 2026.
- [Most MCP memory servers store text. Few tag where it came from.](https://modelbrain.net/articles/mcp-memory-servers-compared.html): Six real MCP memory servers compared on storage, provenance tagging, and account requirements.
- [Local-first software examples](https://modelbrain.net/articles/local-first-examples.html): Eight real examples, Obsidian to Huly, and where the pattern usually stops short of covering an assistant's actual memory.
- [What local-first software actually rules out](https://modelbrain.net/articles/local-first-rules-out.html): A checkable list of guarantees, not a vibe, including the deliberate exceptions (paid sync and backup, Teams, the free ChatGPT connection).
- [ModelBrain's MCP server tools](https://modelbrain.net/articles/five-calls-not-six.html): The original five-call tool surface (September 2026; the server now exposes fourteen tools, listed on the docs page) and why a sixth wasn't added.
- [How ModelBrain tags what it remembers](https://modelbrain.net/articles/how-we-tag-memory.html): Provenance tiers (source evidence, stated assertion, model-generated summary), kept distinct.
- [What a forget call actually has to guarantee](https://modelbrain.net/articles/forget-guarantee.html): Why forget takes an explicit mode (revert, chain, object) instead of one undifferentiated delete.
- [Why we didn't build vault encryption first](https://modelbrain.net/articles/vault-encryption.html): why the local vault relies on OS file permissions and disk encryption, and why we decided (27 Oct) not to add an app-level vault key at all.
- [Designing the recall test before writing the product](https://modelbrain.net/articles/designing-recall-test.html): Five concrete, adversarial questions written down before more of the product, instead of a demo video.
- [What per-node pricing means](https://modelbrain.net/articles/per-node-pricing.html): Billing the machines that talk to each other, not the humans behind them.
- [Why managed backup is included, not billed separately](https://modelbrain.net/articles/backup-is-included.html): Folded into the per-node fee instead of its own metered line.
- [Five questions to ask any AI memory tool](https://modelbrain.net/articles/five-questions.html): A vendor-neutral checklist for evaluating recall claims, including ModelBrain's own.

## Optional

- [Terms of use](https://modelbrain.net/legal/terms.html): The agreement for the free software and the paid services: prices, cancellation, deletion and liability.
- [Privacy policy](https://modelbrain.net/legal/privacy-policy.html): What is held in each mode, retention, processors and your rights.
- [Press](https://modelbrain.net/press.html): The October 11, 2026 release (paid device sync with managed backup and team plans), key facts and press contact, with the October 6, 2026 release archived below it.
- [About](https://modelbrain.net/about.html): Who builds it, and how to reach us.
- [Support & feedback](https://modelbrain.net/support.html): Report a bug, ask for help, or send feedback through a form instead of email.
- [Get it](https://modelbrain.net/get-it.html): macOS (Apple Silicon), Windows (x64 and ARM64) and Linux (x86_64 and aarch64, command line) are available now. macOS is notarized and signed; the Windows installers are signed by NexSpark LLC.