The cheapest way to make Claude actually remember your work in 2026 is not a paid plan or a fancy plugin. It is a folder of plain Markdown notes that Claude can read and write to. Pair Claude (free tier works) with Obsidian (free for personal use), point Claude at the vault folder, and every new chat starts with full context — your projects, your people, the decisions you made last week — instead of the usual blank stare. The whole stack costs $0 if you skip the optional cloud add-ons, and the setup genuinely takes an afternoon.
This guide shows the two connection methods that work on the free Claude tier, the exact structure Claude can navigate without confusion, the limits that will bite you if you ignore them, and the auto-capture layer (an open-source project called Omi) that turns this from a productivity toy into a system that runs itself.
TL;DR — the verdict
Last verified: 2026-07-30. Prices and tier features change — re-check before you commit.
- Best free setup for most people: Obsidian (free core app) + Claude Free tier + the filesystem MCP server. About an hour to wire up.
- No paid plan required. Local MCP servers work on Claude Desktop on any plan, including free. You only need Pro ($20/mo) if you also want Claude Code (the terminal agent) or Opus.
- Two real connection paths: an MCP server (Claude reads/writes/searches your vault through a defined API) or a filesystem connector (Claude treats the vault folder like any other local folder). Pick MCP if you want search and metadata. Pick filesystem if MCP sounds intimidating.
- The honest limit: the Claude Free plan gives you roughly 15–40 messages per 5-hour window (AIToolsRecap, Apr 2026), no Claude Code, and no Opus. Anyone using AI for paid work full-time will outgrow the free tier inside a week.
Why your AI assistant keeps forgetting you
The bottleneck in 2026 is not the model. Claude Sonnet 4.6 (the free-tier model as of July 2026) and Opus 4.8 both have a 1-million-token context window (Claude pricing page). The model is smart enough. The problem is that it has no idea who you are.
Claude's built-in memory helps a little — as of March 2, 2026, all users (including the free tier) get memory across conversations, visible and editable in Settings → Capabilities (AIToolsRecap, Apr 2026). It can remember a name, a preference, a project title. But a single memory line is not a knowledge base. The moment you start a new session it cannot reconstruct your goals, your client list, the decision you made last Tuesday, or the tool stack you standardized on. You spend the first five minutes of every conversation re-explaining yourself. Multiply that across every agent you run and the wasted time compounds.
The fix is to give the model a clean, structured brain it can read from directly. That brain is the Obsidian vault.
What makes Obsidian a good "second brain" for an AI?
Obsidian at its core is a folder of plain Markdown files on your machine. It does not lock your notes in a proprietary format — the file you see in Obsidian is the same .md file any text editor can open, and any AI agent can read. That portability is what makes it different from Notion, Evernote, or Apple Notes for this use case.
Five things matter for the AI-loop specifically:
- You own the files. Obsidian is local-first and your vault is a regular directory. If Obsidian shuts down tomorrow, your notes are still usable in any text editor (Obsidian pricing page).
- No account required for personal use. You can run it with no login and no telemetry. ("No tracking, no account required, no ads" — Obsidian pricing FAQ, verified Jul 2026.)
- The free tier is the full app. Bidirectional links, graph view, Canvas, and 4,000+ community plugins are free, with no usage caps on a single device (Obsidian pricing page).
- Markdown is the format AI expects. Plain headings, bullet lists, tags — Claude and every other LLM parses Markdown natively. There is no extraction step.
- Backlinks and tags give the model a map.
[[note]]wikilinks and#tagmarkers let Claude walk your knowledge graph the way a human would, not as opaque binary blobs.
For this article I am assuming personal use. Commercial use is technically covered by the $50/user/year license, but as of February 2025 Obsidian made it voluntary — using Obsidian at work is free, organizations just buy the license to support development (Obsidian EULA update, Feb 2025). Paid add-ons you might want later: Obsidian Sync for end-to-end encrypted cross-device sync at $4/user/month billed annually, and Obsidian Publish to host your vault as a website at $8/site/month billed annually (Obsidian pricing page). Neither is required for the Claude loop — they are convenience, not dependencies.
How to connect Claude to an Obsidian vault — both methods
There are two genuinely free paths. Pick one — you do not need both.
Method 1: The filesystem MCP server (the easiest on-ramp)
If "MCP" sounds scary, start here. The filesystem server does exactly what it sounds like — it hands Claude read and write access to a folder on your machine. Point it at your vault directory and Claude can open files, list them, edit them, and create new ones in plain English.
What you need: Claude Desktop (the desktop app, free) and Node.js installed on your machine. You do not need a Pro subscription for local MCP servers — local MCP works on any plan, including Free (Claude features reference, Suprmind, updated Jul 2026).
Step-by-step:
- Install Obsidian from obsidian.md (free, no account). Create an empty vault in a folder you know the absolute path to, e.g.
~/Documents/vault. - Install Claude Desktop from claude.com/download (free).
- Open the Claude Desktop config file:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json - Linux:
~/.config/Claude/claude_desktop_config.json
- macOS:
- Add the official
filesystemMCP server (published by the Model Context Protocol org), pointing at your vault path:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents/vault"]
}
}
}
References: MCP.filesystem reference server (ModelContextProtocol org); Claude Desktop MCP setup docs.
- Restart Claude Desktop. You should now see filesystem tools in your chat. Ask Claude to "list the files in the vault" — it will be able to read them.
That is the whole setup. You can now ask Claude, in plain language:
- "Read every note tagged #marketing and give me a one-paragraph summary of each."
- "Create a new note called
2026-Q3-customers.mdand put these three contacts in it." - "Find every file mentioning 'Mercury' and tell me where it is referenced."
The vault folder is treated like any other local directory — Claude just opens files as needed.
Method 2: An Obsidian-specific MCP server (richer, slightly more setup)
If you want Claude to respect Obsidian's own features — wikilinks like [[note]], frontmatter, the tag index, the metadata cache — use a community Obsidian MCP server. The trade-off is one more layer to configure.
Two well-supported options:
@istrejo/obsidian-mcp— an Obsidian MCP server with 13 tools for reading, writing, searching, and organizing notes. Direct filesystem access, no Obsidian plugin required, works even if Obsidian is closed. Works with both Claude Desktop and Claude Code over stdio. MIT-licensed. (GitHub: istrejo/obsidian-mcp)bdwalter/Obsidian-mcp— an Obsidian community plugin that runs the MCP server inside Obsidian itself and exposes the vault over Streamable HTTP. This one gives Claude access through Obsidian's own API, so it sees backlinks, frontmatter, the periodic notes index, and the tag index that the filesystem server does not. (GitHub: bdwalter/Obsidian-mcp)
Setup is the same pattern: add the server to claude_desktop_config.json, restart Claude Desktop, verify. The Obsidian-specific servers are better when your vault is large and searchability matters more than absolute simplicity.
How do the two methods compare?
| Method | What Claude sees | Best for | Setup difficulty | Cost |
|---|---|---|---|---|
| Filesystem MCP server (Anthropic/MCP org) | Files in the folder — opens them like a text editor | Beginners; "just point me at a folder" | Low — one config entry | Free |
@istrejo/obsidian-mcp |
Files + search + organize tools, but treats vault as a directory | Mid-size vaults where you want Claude to search and reorganize | Medium | Free |
bdwalter/Obsidian-mcp (runs inside Obsidian) |
Builds on Obsidian's metadata — wikilinks, frontmatter, tag index, backlinks | Power users; large, interlinked vaults | Medium-high | Free |
A note on the Claude Desktop side: local MCP server support works on every plan, but remote MCP connectors (Anthropic-hosted, no install) are gated — those need Pro, Max, Team, or Enterprise (Claude features reference, Suprmind, updated Jul 2026). For the Obsidian use case you do not need remote connectors — the entire vault lives on your machine.
Structure: the four-bucket system Claude understands
Now the file system works. The next problem is what to put in it. Dump five years of notes in a flat folder and the AI gets lost the same way you would.
PARA — Tiago Forte's Projects / Areas / Resources / Archives structure — is the natural fit. Claude already understands PARA because it appears everywhere in its training data; you do not have to teach it the rules. You tell it "this is a PARA vault" in one prompt and it can sort, link, and reorganize.
What is the PARA method?
PARA was first published by Tiago Forte in a 2017 blog post and later expanded into The PARA Method book in 2023, building on his 2022 Building a Second Brain (Fabric, "The PARA Method: A Complete Guide," Jul 2026). The four categories are:
- Projects — short-term efforts with a clear goal and a deadline. A launch, a course, a client deliverable. Active work you are pushing on now.
- Areas — ongoing responsibilities with no end date. Marketing, finances, a client relationship, your health. Domains you maintain.
- Resources — topics of ongoing interest for future reference. Reference material, articles worth saving, books, frameworks worth keeping. Not actionable today, but useful later.
- Archives — inactive items from any of the other three categories. Completed projects, abandoned resources, former areas. Still searchable, out of your way.
The organizing principle is actionability, not subject. The same PDF about marketing strategy can be a project deliverable, an area reference, a general resource, or an archived artifact — where it belongs depends on its role in your life right now, not its topic. That distinction is what makes PARA survive scale when topic-based folders collapse.
How to set up PARA in Obsidian (numbered steps)
- Open your vault and create four top-level folders:
1-Projects,2-Areas,3-Resources,4-Archives. Number them so they sort in this order. - Inside
1-Projects, create one folder per currently active project. Tiago Forte recommends 10–15 active projects maximum — if you have more, archive some (Forte via Fabric, Jul 2026). - Inside
2-Areas, create a folder per ongoing responsibility. - Do not pre-create empty resource folders. Create a
3-Resourcesfolder only when you actually have something to file (Forte via Fabric, Jul 2026). - Create one empty
4-Archivesfolder. Forte's rule: never delete, always archive. Storage is cheap and you cannot predict what will turn out to be useful later. - Add a top-level note called
0-Index.mdlisting every active project and area in a one-line summary. Claude can read this to orient itself cheaply — much faster than reading every file. - Add an
AGENTS.mdfile (orCLAUDE.md) at the vault root explaining your rules: "When you receive a new note, save it raw, then suggest tags and a PARA bucket." Claude Desktop and Claude Code both read files like this on every session.
How does information move through PARA?
PARA is not a filing cabinet. Information flows between buckets as your life and work change. A Resources interest in machine learning generates a Projects goal to finish an ML course by June. The project finishes and moves to Archives, but your interest in ML stays in Resources. An Areas responsibility like career development spawns multiple projects over time — update CV, attend conference, get certified — each one finishes and archives. When an area is no longer your responsibility, the whole thing moves to Archives. This flow is why a topic-based system eventually breaks and PARA does not.
A worked example — what Claude actually looks at
Imagine you ask Claude to "organize yesterday's calls and tell me what to follow up on." Here is roughly what happens:
- Claude reads
0-Index.mdto orient itself. - It reads
AGENTS.mdto learn your rules. - It looks in
1-Projects/for any active project that touches what you mentioned. - It writes raw notes into the right project folder, tagged and backlinked.
- It extracts action items into a
tasks.mdfile in that project.
That five-step loop runs without you opening a single folder. The structure does the work.
Adding the capture layer: why Omi makes this run itself
So far this is a manual setup — you write notes, Claude reads them. The next jump is to make capture automatic.
There is an open-source project called Omi (GitHub: BasedHardware/Omi) that is the missing piece. Omi is an ambient-AI app — originally paired with an open-source wearable but usable as a phone app — that records your conversations, transcribes them, identifies tasks, and builds a local knowledge graph from your day. It is fully open-source (Hardware + Software) and the core app costs nothing. The "Omi chat" subscription (an AI inside the Omi app) is the only paid piece, and you can skip it without losing capture (Omi GitHub org).
For the Obsidian loop specifically, Omi's value is that it writes finished Markdown summaries — not raw transcripts — that automatically sync into your vault. You do not take notes; Omi captures your day, and Claude reads the result.
A pragmatic flow looks like this:
- Omi captures the day: meetings, voice memos, phone calls (with consent).
- Omi writes a daily summary file into your
3-Resources/Journal/folder or2-Areas/Daily/folder in your Obsidian vault. - Obsidian shows them as a graph, tagged and interlinked.
- Claude reads the vault every session, understands what you did yesterday and what's current today, and can reorganize PARA buckets as you go.
The video version of this exact stack talks about an Obsidian vault with around 1,800 memories auto-pulled in. For someone who takes 20–30 calls a month, that compounding context is the upgrade from "neat tool" to "the AI genuinely knows my business."
The free-tier limits nobody mentions
The setup above works on the Claude Free plan — but the Free plan has real ceilings. Here is what you should know before you commit to it, all verified against Anthropic's pricing page and a tier-by-tier breakdown:
What is included free in 2026?
Per AIToolsRecap's free-tier breakdown (Apr 2026) and Anthropic's own pricing page:
- Claude Sonnet (the latest current version; Sonnet 4.6 as of mid-2026) — the model itself, no card required.
- Web search inside chats.
- Memory across conversations (since March 2, 2026), viewable and editable in Settings.
- Up to 5 Projects — each is a self-contained workspace with custom instructions and knowledge base files.
- Artifacts, extended thinking, and code execution.
- Local MCP servers, including the filesystem and Obsidian servers above.
What is NOT free?
- Claude Code — the terminal-based agentic coding tool. Requires Pro ($20/month) or API credits. The free tier's chat can answer coding questions and write code, but the autonomous, multi-file, codebase-aware terminal agent is paid.
- Opus — the frontier reasoning model. Sonnet only on Free.
- Claude Research — the deep multi-source synthesis mode. Pro-only.
- Cowork (computer-use and background tasks). Pro and Max only.
- More than 5 Projects. Free is capped at five.
- No Claude Code, no Opus, no multi-source research features (AIToolsRecap, Apr 2026).
How many messages can you actually send?
Free-tier users get roughly 15–40 messages per 5-hour window depending on how many long documents you upload and how much thinking each one triggers (AIToolsRecap, Apr 2026). There is no published hard cap — Anthropic uses a dynamic limit tuned for fairness. The honest signal to upgrade is not hitting the wall once; it is hitting it consistently enough that the interruptions break your workflow.
Practical implication for the second-brain setup: if you point Claude at a 50-MB vault and ask it to summarize everything every chat, you will blow the limit fast. The workaround is 0-Index.md and selective reading — Claude reads one small index file, not the entire vault, so each turn burns a small fraction of the budget.
What this means for you
If you are using AI for your work — running a small business, building a product, managing clients — the cost-benefit on this stack is unusually clear:
- Under 30 min a day of AI use: Stay on the Claude Free plan. Wire up the filesystem MCP server against your Obsidian vault, use PARA, and lean on the digest/index pattern. You will rarely hit rate limits and pay nothing.
- 1–3 hours a day, mission-critical: Upgrade to Claude Pro ($20/mo). The 5× message capacity is meaningful, plus you get Claude Code (which is a second brain of its own — it can read and modify your vault through the same MCP server) and Opus when you need it.
- Multiple agents in parallel (e.g. a Hermes fleet, a team using Claude Code simultaneously): You are no longer in the "free stack" mindset. Treat the Obsidian vault as shared state each agent reads from and writes to, and see our deeper write-up on building an agent OS with Hermes and Obsidian.
The architectural lesson is the same at any tier: the model is not the bottleneck, the context is. Every dollar you spend on a smarter model is wasted if it cannot see your work. A free Claude pointed at a structured vault beats a paid Claude pointed at nothing — every time, on every query.
For more on context engineering as a discipline (which is what this really is), see our deeper guide to auditing your prompts for Claude 5 and the skill-by-demonstration replacement for classic prompt engineering. If you want a turn-key alternative that does not require setting up Obsidian, Genspark SecondBrain gives you a hosted persistent-memory layer out of the box, and tools like Gemini Notebook collections cover the same idea for a research-and-reading workflow.
The backup rule before you let AI reorganize anything
One hard caveat: the moment you let Claude restructure a vault — move notes between PARA buckets, batch-rename, batch-tag — you have handed an AI write access to your knowledge base. That is exactly what you want, except when it goes wrong.
Before any batch change:
- Back up the vault. The simplest method is Git (
git initin the vault root, commit, and you cangit reset --hardto undo anything). If you do not know Git, copy the folder. - Start small. Test on one folder before running a "reorganize everything" prompt.
- Keep raw notes raw. Do not overwrite human notes with AI rewrites — keep the source and let Claude add a synthesis below or in a separate file. Long term, this preserves provenance and lets you trust the vault.
The backup rule is not optional. The Obsidian Sync add-on ($4/mo billed annually) also gives you version history, which is a reasonable belt-and-suspenders option if Git sounds like too much.
FAQ
Q: Does the free Claude plan actually work with Obsidian? A: Yes. Local MCP servers — including the filesystem server that points at your vault — work on Claude Desktop on any plan, free included. You do not need Pro unless you also want Claude Code (the terminal agent), Opus, the Research mode, or more than five Projects. Verified against Anthropic's pricing page and Suprmind's Claude features reference (updated Jul 2026).
Q: Is Obsidian really free? A: The core app is free forever for personal use, with no feature caps on a single device and no account required. Paid add-ons are Sync ($4/user/month billed annually) and Publish ($8/site/month billed annually). A $50/user/year Commercial License is voluntary since February 2025 — using Obsidian at work is otherwise free. Source: Obsidian official site and Obsidian pricing FAQ.
Q: What does PARA stand for? A: Projects, Areas, Resources, Archives — the four categories created by Tiago Forte in a 2017 blog post and turned into a 2023 book, The PARA Method. The organizing principle is actionability, not topic: Projects are most actionable, Archives least, and the same document can move between them as your work changes. Source: Fabric, "The PARA Method: A Complete Guide" (Jul 2026).
Q: Can Claude modify my Obsidian vault, or only read it? A: Both. With a read-write MCP server (filesystem, Obsidian-specific MCP servers), Claude can read, search, create new notes, reorganize folders, and edit existing files. You can also restrict any of these servers to read-only mode if you want to keep a wall between reading and writing.
Q: How much does the Omi capture layer cost? A: The Omi app and the entire Omi platform are open-source (MIT licensed) and free. The optional "Omi chat" subscription (an AI inside the Omi app) is the only paid piece — and you can skip it without losing the capture-to-Markdown loop. The complete source is at github.com/BasedHardware/Omi.
Q: What breaks if I upgrade to Claude Pro ($20/mo)? A: Nothing breaks — Pro is a strict superset. Pro gives you roughly 5× the message capacity of Free, Opus access, Claude Code (the terminal agent), the Research mode, unlimited Projects, and priority access during peak hours. For a second-brain workflow that you use for more than 30 minutes a day, Pro pays for itself inside a week.
Discussion
0 comments