Obsidian AI Agent Memory: The Daily Log Setup

A shared Obsidian AI agent memory lets every agent read what the others did yesterday and pick up where they left off. Each agent writes one nightly summary into a single vault note, and one orchestrator agent sets the whole thing up from a single prompt. This guide walks through the folder structure, the prompt, and the daily log format.

How the Obsidian AI Agent Memory Daily Log System Works

The agent daily log system is a shared vault in Obsidian, a note app that stores everything as plain Markdown files on your local disk. Every night, each agent writes one summary of its day into the vault: what it worked on, what it finished, what it learned, decisions made, and what is still in progress. The next morning, any agent reads that note in under a minute and continues where the others stopped.

One correction to the video's framing: Obsidian is free for personal use but proprietary software, not open source, according to Obsidian's own site. The memory itself is open, though: plain Markdown files you can read, back up, or move to any other tool. The automation does not come from Obsidian at all. Your agents read and write the files directly, so nothing needs a plugin or a paid sync service.

The system is agent-agnostic. The video demonstrates it with Hermes Agent alongside Claude Code, OpenAI's Codex CLI, and OpenClaw, an open-source personal AI assistant. Any agent that can read and write files on your machine can join the same vault.

The Vault Folder Structure: PARA Plus an Agent OS

The vault uses the PARA method, a four-folder organization scheme documented by Tiago Forte in 2017: Projects, Areas, Resources, and Archive. On top of PARA, the video adds four folders that make the vault usable by agents: an inbox, a daily notes folder, a memories folder, and an Agent OS folder.

Each folder has a specific job:

• Inbox holds messy captures. An agent drops something here when it is unsure where it belongs.
• Daily holds your own daily note, such as top three priorities and notes for tomorrow.
• Projects holds active work, one folder per project, each with an index.md.
• Areas holds ongoing responsibilities, such as SEO, community, or content.
• Resources holds reference material: about-me pages, SOPs, staff notes, and content writing assets.
• Memories holds decisions, the agent log, and wins.
• Archive holds finished or inactive material.
• Agent OS holds the protocol and per-agent instruction files so everything syncs from one place.

The Agent OS folder is what turns separate tools into one memory galaxy. In the video's setup, notes link across folders, so a note about a person links to projects, funnel strategy, and content written for them. Because agents maintain the links, you never organize the vault by hand.

One Mega Prompt and an Orchestrator Agent

Setup runs through a single prompt given to one orchestrator agent rather than configuring each tool separately. The video recommends Claude Code or Codex as the orchestrator because both can read and write files across your machine, which is the only capability the setup needs.

The orchestrator performs five steps from that one prompt:

1. Build the folder structure at the vault path you specify.
2. Write the shared protocol as one note inside the vault.
3. Add a short pointer to each agent's own instruction file, telling it the vault exists and to read it at the start of every session.
4. Add a nightly task so each agent writes its summary without being asked.
5. Run a test task for each agent and report the results.

Specify exactly which agents to set up, for example Claude Code, Hermes, Codex, and OpenClaw. The video reports that the same prompt also works on an existing vault: the orchestrator inspects what is already installed and figures out what is synced and what is missing. The speaker's demonstration is first-hand experience, not an independent benchmark, so expect to review the instruction files it writes before trusting the schedule.

What Each Agent Receives After Setup

After the orchestrator finishes, exactly two things land in each agent. The first is a short pointer in its instruction file so it knows the vault exists and reads it at the start of each session. The second is a scheduled nightly task that writes the daily summary on its own.

The rules themselves live in one note inside the vault, not duplicated per tool. This matters when you change anything: you edit one protocol note, and every agent reading the vault picks up the new rule. The video shows the orchestrator working out per-agent details automatically, including which instruction file each tool reads, where it keeps history, and how it schedules tasks.

The Daily Log Format and Useful Extra Prompts

Each agent writes one summary per day, not one entry per task. The video's format has five fields: worked on, learned, decisions, in progress, and finished. The learned field is the one that compounds. A self-improvement captured inside one agent benefits only that agent; written into the shared vault, it benefits every agent that reads the log the next morning.

Four optional prompts extend the basic loop. Use one or two, not all of them:

• Morning brief: read yesterday's logs and send a five-bullet brief covering what is done, what stuck, and the top three priorities. The prompt explicitly forbids writing to the vault, keeping the brief read-only.
• Pre-task check: before starting, read the last three days of Agent OS memories for anything about the current project, report what is relevant, then continue.
• Manual log: trigger today's summary on demand instead of waiting for the nightly run.
• Health check: ask the orchestrator to list each agent's scheduled task, when it last ran, and whether yesterday's log has its own section, then fix any failed task and report the changes.

A weekly roll-up is another option the video mentions. These extras are ideas rather than a required stack; the daily log alone is the recommended core.

Why Millions of Memories Break This System

Do not dump full conversation histories into the vault. The video's warning is about token cost: every memory an agent reads at session start consumes context, and a vault with unbounded entries becomes too large and too expensive to load. Community members in the video asked how to store millions of memories in Obsidian, and the speaker's answer is that this is the wrong tool for that scale.

A vector database can handle that volume, but the video argues it is unnecessarily technical for most individual users. The daily log design is the middle path: one concise, structured summary per agent per day. Ten agents produce ten short entries per day, and the whole vault's recent history stays small enough for any agent to read in under a minute. If your use case genuinely needs million-scale recall, plan a vector store from the start instead of outgrowing this setup.

Variations, Capture Tools, and Content Workflows

The daily log is the recommended option, but the video lists alternatives for people who want to experiment: an entry after every task, one file per agent per day, agents writing into your own daily note, agents filing straight into PARA, and an auto-generated wiki of people, tools, and topics.

Automatic capture tools go further. Omi, an open-source AI wearable, and screenpipe, which records screen and microphone activity into a searchable local history, can turn what you do and say into memories that sync into the same vault. The speaker used capture tools in the past and reports agents are now good enough that he no longer needs them.

One workflow built on top of the memory: a create section that reads recent vault entries and generates an app, SEO article, social post, or script from them. Because content creation is usually documenting real work, a vault that already holds your recent projects becomes raw material for that content without extra research.

Frequently Asked Questions

• Which agents can share one Obsidian vault? Any agent that can read and write local files can join. The video demonstrates Claude Code, Codex, OpenClaw, and Hermes Agent in one vault. The orchestrator writes a per-agent pointer into each tool's own instruction file, so each tool reads the same protocol note its own way.

• Do I have to edit vault files myself? No. One orchestrator agent, such as Claude Code or Codex, builds the folders, writes the protocol, and schedules the nightly tasks from a single prompt. The daily summaries are also written by the agents, not by you.

• Is Obsidian free? Yes, Obsidian is free for personal use, but it is proprietary software rather than open source, as its official site states. Your memory data itself is plain Markdown stored locally, so it remains portable to any other tool.

• Should I store full agent conversations in the vault? No. Full histories inflate the vault and the token cost every time an agent reads it. The daily log format of one concise summary per agent per day keeps the whole shared memory small enough to read in under a minute.

Source video