Frequently Asked Questions.

Agent Memory is a local-first semantic memory app for AI agents. You install it once on your machine, point it at folders of project knowledge — Markdown docs, source code, runbooks, decision logs, manual notes — and it indexes them locally with the open-source Xenova/all-MiniLM-L6-v2 embedding model.

Once running, it exposes a small HTTP API on 127.0.0.1:8765 that any AI assistant on your machine can call. The same memory shows up in every agent — no re-explaining your project at the start of every conversation. Eighteen ready-made Skills across Organising, Design and Marketing ship in the box.

Three big differences:

• It runs on your machine. Your files, embeddings and search index stay local. No prompt is sent to a third party unless your assistant chooses to relay a snippet as part of the conversation, under their privacy terms.
• It works across agents. ChatGPT Memory only helps inside ChatGPT. Claude Projects only helps inside Claude. Agent Memory is a single store every assistant on your machine can read from.
• It indexes raw folders. You don’t have to upload, paste or curate. Point it at a repo or a docs folder and it walks the tree.

Agent Memory is a one-time $29 USD purchase, with a lifetime licence. There is no required subscription. An optional Updates Pass at $5 / month adds new skills and feature updates as they ship — cancel any time and the software keeps working.

Cloud assistants you connect (Claude, ChatGPT, etc.) keep their own pricing — you pay them directly. See Pricing for the full breakdown.

Agent Memory runs on Windows, Mac and Linux. The minimum is roughly 4 GB RAM and a few hundred MB of free disk for the model cache and index. The default MiniLM embedding model is small enough to run comfortably on a CPU; no GPU is required.

For platform-specific details see the Windows, Mac and Linux entries below.

A skill is a small reusable recipe — a Markdown + JSON file — that says what to read from local memory, what to produce, and how the assistant should style its reply. Skills live alongside your memory store. Any connected assistant can call one by name (“Run Project Briefing for X”), and Agent Memory hands back the right snippets and your profilePrompt in one go.

Eighteen, across three categories — six per category:

• Organising — Project Briefing, Decision Recap, Runbook Writer, Daily Standup, Risk Spotter, Onboarder.
• Design — Brand Guardian, UX Reviewer, Pattern Library, Asset Tracker, Reading Order Audit, Style Keeper.
• Marketing — Site Auditor, Launch Brief, Voice Match, Press Kit, FAQ Builder, Social Snippets.

Every skill is editable. You can copy one, write your own, or share a folder of skills across machines.

Yes — new official skills are added through the optional Updates Pass. They drop into your skills folder automatically, you keep using them forever, and you can edit them in place like any other skill.

Absolutely. A skill is a tiny Markdown + JSON file that names the folders to read, the file glob, the output format, and the assistant style. Agent Memory loads the folder on startup and exposes every skill through the same local API. The home page has a worked example.

Windows 11 64-bit (recommended) and Windows 10 64-bit version 22H2 or later.

Two builds are available. The full Electron build ships as Agent Memory Setup x.y.z.exe. The lightweight build ships as AgentMemory-Light-Windows.zip and uses the Edge browser already on the machine instead of bundling Chromium — substantially smaller download.

You extract the .zip, then double-click Double Click to Install.exe. The launcher shows the legal agreement, installs runtime dependencies on first launch, starts the local API and opens Agent Memory in an Edge app window. Logs are available from View AgentMemory Logs.cmd in the same folder.

The lightweight build currently expects Node.js 22+ to be installed on the machine. A future installer will bundle a service runtime so this requirement goes away.

The JSON index, sources list, user guidance profile and skills folder live under %APPDATA%\AgentMemory\memory-store. The Edge profile used by the lightweight build lives under %LOCALAPPDATA%\AgentMemory\EdgeProfile. Back them up like any other local files.

macOS 13 Ventura or later, on Apple Silicon (M1 / M2 / M3 / M4) or Intel x86_64. The Electron build ships as a .dmg and a .zip for both architectures.

Minimum: 8 GB unified memory, 500 MB free disk for model cache and index.

Recommended: 16 GB+ unified memory and an SSD with 2 GB+ free, especially if you index large source repositories.

The Electron build stores data under ~/Library/Application Support/Agent Memory/memory-store. Uninstall is a drag-to-trash on the app, plus removing that support folder if you want a clean wipe.

Any modern x86_64 distribution. The Electron build ships as an .AppImage, a .deb for Debian/Ubuntu, and a .tar.gz. Tested on Ubuntu 22.04 / 24.04 LTS, Debian 12, Fedora 39+, Arch (rolling), Pop!_OS and Linux Mint 21+.

Minimum: 4 GB RAM, 500 MB free disk, a 64-bit x86 CPU from the last several years.

Recommended: 8 GB+ RAM and an SSD if you index multi-gigabyte repositories.

User data lives under ~/.config/Agent Memory/memory-store (or the equivalent $XDG_CONFIG_HOME path). Back it up like any other dotfile.

No, by default. Indexing, embeddings and search all run locally. The first time the app runs, the Xenova/all-MiniLM-L6-v2 model is downloaded once from Hugging Face and then cached locally. After that, you can work fully offline.

Agent Memory itself does not send your file contents, embeddings, queries or guidance profile to any server. The local API binds to 127.0.0.1, not your network interface, so other machines on your network cannot reach it without you explicitly opening a port.

If you point a hosted assistant — Claude Desktop, ChatGPT Desktop, Cursor — at Agent Memory and that assistant relays a search snippet to its provider as part of the conversation, that part of the conversation goes under the assistant’s privacy terms. Agent Memory does not relay anything for them.

If privacy matters more than convenience, run a fully local assistant (for example a local LLM via Ollama or LM Studio) and have it call Agent Memory’s local API. Nothing leaves your machine.

Agent Memory walks the folders you choose and indexes supported text-like file types up to 2 MB each. Today that includes Markdown, plain text, source code (TypeScript, JavaScript, Python, Go, Rust, Java, C#, PHP, Ruby, Swift, Kotlin, etc.), shell scripts, configuration formats (JSON, JSONL, YAML, TOML, CSV) and HTML/CSS/SQL/XML.

Folders called node_modules, .git, dist, dist-installer and dist-edge are skipped automatically.

Agent Memory does not send telemetry of conversation content, file content, indexed paths, or queries. There is no analytics SDK in the desktop app. Updates and model downloads are explicit operations.

The default is Xenova/all-MiniLM-L6-v2, an ONNX conversion of sentence-transformers/all-MiniLM-L6-v2. The declared model licence is Apache-2.0. It’s small (~25 MB), fast on CPU, and produces good general-purpose semantic embeddings for source code and docs.

The model name is stored in the local index and used by the embedding pipeline. Power users can edit the JSON store to switch to another model supported by @xenova/transformers, then re-index. A first-class “choose model” UI is on the roadmap and will roll out via the Updates Pass.

No. Agent Memory only does embeddings and search. The LLM lives in whichever assistant you connect — Claude, ChatGPT, Cursor, or a local model via Ollama or LM Studio. Agent Memory is the memory layer, not the chat model.

Any tool that can speak HTTP can use Agent Memory. The recommended pattern is to call POST http://127.0.0.1:8765/search at the start of a task, treat the returned profilePrompt as system context, then read the ranked snippets.

POST http://127.0.0.1:8765/search
Content-Type: application/json

{ "query": "What did we decide about pdfaa.ai staging?",
  "limit": 8 }

See the full connection guide for copy-paste examples for Claude Desktop, ChatGPT Desktop, Cursor, n8n and your own scripts.

Anything that can speak HTTP. We have ready-made instructions for Claude Desktop, ChatGPT Desktop, Cursor, Continue.dev, n8n, Zapier and plain shell scripts. Feature-parity comes from the agent — Agent Memory exposes the same API to every caller.

It’s a small system-prompt-style preface built from your About Me, How I Like to Work and Very Important answers. Every /search response includes it before the ranked results, so assistants apply your working preferences consistently.

$29 USD, one-time, for a lifetime licence. You install the app, you own that version forever.

An optional $5 / month subscription that delivers ongoing updates: new skills, feature improvements, refreshed embedding-model picks, and access to the Agent Memory community channel. Cancel any time — the version of Agent Memory you bought keeps working.

Yes — 30-day money-back, no questions asked. Email hello@agentmemory.cc from the address you bought with and we’ll refund you.

Yes. Site-wide and reseller licences are available. Email hello@agentmemory.cc with the rough team size and we’ll send terms.

Agent Memory is proprietary commercial software from so.design. It statically links and ships permissively-licensed open-source components (Express MIT, CORS MIT, @xenova/transformers Apache-2.0, MiniLM Apache-2.0). Full attributions are on the Attributions page.

Yes, under the Agent Memory Commercial Licence. Redistribution, resale, hosting, white-labelling, reverse engineering, removal of notices, bundling with another product, or use to build a competing product requires written permission from so.design. See Licensing.

Full Terms of Use, Commercial Licence and Attributions are on the legal page. The desktop installer also requires explicit acceptance of all three on first launch.

The Electron desktop builds (Mac, Linux, full Windows) ship everything as native binaries — no runtimes to install. The current lightweight Windows build expects Node.js 22+ on the host machine. A future bundled installer will remove that requirement.

Yes. The index is a single JSON file (index.json) under your user data folder. It contains the model name, sources you’ve added, the JSON-encoded chunks, and their embeddings. You can read it, back it up, or delete it to start fresh.

Agent Memory automatically tries the next available local port (up to a small range above 8765) and shows the exact address in the app header. Agents should read the active port from GET /stats rather than hard-coding it.

Email hello@agentmemory.cc. Reproducible repro steps, log output (from View AgentMemory Logs.cmd on Windows or the terminal on Mac/Linux) and the version from GET /stats are very welcome.