Your knowledge.
Any agent.
Built to last.
One memory every AI tool can read and write, in Markdown you own. Browse it as a wiki, build live notes on it, and put agents to work keeping it current.
A decision note: what was decided, why, and what it connects to. It's a Markdown file, so you and every agent read the same note.
Your knowledge, as a workspace.
Kept current by agents.
Open a project like a wiki.
Not a folder of files.
Every project gets an overview page that links into its decisions, research, meetings, and people. Anyone on the team can find their way around, and every page is still a Markdown note your agents read and write.
- Every folder keeps an index.md and a log.md in the Open Knowledge Format, updated as notes change
- Folders, notes, and an outline for every page
- The same notes your AI tools use, no copy to keep in sync
Notes that show the work.
Boards and charts, written in Markdown.
Basic Memory supports Comark, an open Markdown syntax for components. Put a KPI, a Gantt chart, a comparison, or a task board inside any note: the data sits in the note as ordinary lists and tables, and Basic Memory draws the figure. Agents write these components the same way you do, and read them back as plain text, numbers included.
- 39 components: boards, charts, KPIs, timelines, calendars, comparisons
- Live views query your notes each time you open them, so a chart of open tasks stays current
- Still plain Markdown: in any other editor, or to any agent, it reads as text
Try every component in the playground ↗About Comark: comark.dev ↗
--- title: Launch status type: report owner: Maya Chen --- # Launch status Week of February 22. Launch is March 12. ::bm-kpi{title="Customers migrated" value="26" label="of 40 in cohort one" hint="65%" data="0,3,7,12,16,21,26"} :: ::bm-funnel{title="Beta switch funnel"} - Saw the new plans: 112 - Opened the usage chart: 84 - **Set a spending limit: 57** - Switched plans: 31 ::
Workstreams and how far along each one is.
Bring the documents in.
Keep the original.
Drop in PDFs, Word files, slides, and spreadsheets. Each becomes a searchable note your agents can read and search, with the original file one click away.
- PDF, Word, PowerPoint, and CSV
- Extracted text keeps headings and tables
- Plus your Claude and ChatGPT history
Agents that keep it current.
And show their work.
Give an agent instructions in a note and let it follow up on meetings, link related notes, or fix facts that went stale. Every run shows exactly what it created and changed, and you can keep it, revert a note, or undo the run.
- Run on demand, or when a new note lands in a folder
- Ready-made agents to start from, or write your own
- Billed by usage credits; failed runs cost nothing
People and agents,
on the same page.
A shared memory only works if both sides can read it, change it, and check each other's work. These are the parts built for that.
Suggest, don't overwrite.
People and agents leave comments and suggested edits right in the note, and anyone with access accepts or rejects them. Review is stored as CriticMarkup inside the Markdown file, so it stays readable plain text that moves with the note.
Your notes, inside the conversation.
Pocketbook, the Basic Memory app for Claude and ChatGPT, opens your notes right in the chat. Browse a project, search it, see how it connects, and add any note to the conversation without switching windows.
Read a note in the conversation, then hand it to the model.
Browser agents, as you
In browsers that support WebMCP, the web app gives browser agents its own tools: search, read, and open notes, suggest edits for you to review, check live components, and download snapshots. They act as the signed-in person, in the workspace that's open.
Live updates and a merge view
Changes from agents and teammates appear without reloading. If you and an agent edit the same note, you see both versions and merge them instead of losing either.
Every change has an author
Notes record which person or agent created and last edited them, and agent edits carry the run that made them.
Any model. Any subscription.
The same memory.
Save a decision in one tool and use it in a fresh session in another. Switch models, switch vendors, run three agents at once: they all work from the same notes, and nobody pastes context between chats.
- Claude
- ChatGPT
- Claude Code
- Codex
- Cursor
- VS Code
- Gemini CLI
- OpenClaw
- Hermes
- Any MCP client
One URL, every device
Hosted memory over MCP. Sign in once with OAuth and use it on web, desktop, and mobile.
Bring your own harness
Plugins for OpenClaw, Hermes, and Pi, ready-made skills, and session hooks any agent harness can call.
Coding sessions that remember
Claude Code and Codex plugins brief each session from memory and checkpoint before context is compacted.
Keys for headless agents
API keys for scripts and autonomous agents, read-only or limited to the projects they should see.
Readable today.
Portable forever.
Your memory is Markdown you can open, edit, and move. This is a whole decision note, exactly as it sits on disk. If you leave, it comes with you, intact.
- Every note is a normal .md file, readable in any editor
- Open the same folder in Obsidian or VS Code, no plugin needed
- Desktop sync keeps a local copy you can edit anywhere
- Export a project as a ZIP anytime
- Open source under AGPL; self-host with SQLite or Postgres
--- title: Harbor launch date type: decision status: decided owner: Maya Chen --- # Harbor launch date We moved the launch from February 26 to March 12. ## Observations - [decision] Launch moves to March 12 #harbor - [reason] Billing migration must finish first - [alternative] Slip to April, rejected ## Relations - decided_in [[Jan 14 planning]] - depends_on [[Billing migration]] - blocks [[Launch checklist]]
Speaks OKF, the open standard for knowledge bundles.
The Open Knowledge Format is an open spec from Google Cloud for Markdown knowledge that any agent can navigate: typed notes, an index.md for each folder, and a log.md of changes. Basic Memory's wiki keeps those files current as your notes change, and bm okf export writes a whole project as an OKF bundle.
- Typed frontmatter on every note; links become standard Markdown links
- Relation types kept in metadata, so nothing about the graph is lost
- Checked against the spec before it's written, with bm okf check for any bundle
harbor-launch/ ├── index.md okf_version: "0.2" ├── log.md changes by date, newest first ├── decisions/ │ ├── index.md what's in this folder │ ├── beta-scope.md │ └── harbor-launch-date.md └── research/ ├── index.md └── plan-comparison.md
Find the decision,
not just the keyword.
Search by meaning
Hybrid semantic and keyword search finds the note that answers the question, even in different words.
Query like a database
Filter on any frontmatter field: status, owner, date, anything you track.
Follow the connections
Typed relations and categorized facts form a graph your AI walks from a decision to the research behind it.
Structure that holds
Define note types with schemas, validate them, infer them from what you have, and catch drift.
Nothing gets lost.
No surprise bills.
Every version kept
File history keeps every version of every note, with compare and restore.
Snapshots and backups
Daily snapshots, manual ones you keep, restore for a file, folder, or project, and a verified daily backup with a second provider.
Agents can't clobber you
Revision-safe writes mean an agent can only replace the version it read, and locked notes are off limits to agents.
Your own isolated workspace
Each workspace gets its own database and storage, not rows in a shared table.
Roles and access
Owner, Admin, Editor, and Viewer, plus invite-only projects, User Admin, and an audit log on Business.
Usage you can see
Credits by day, person, app, API key, and agent, with a spending limit on from day one.
The full list.
Notes, search, and the tools for your AI are open source and run on your own machine. Cloud adds the web app, teams, history, and agents.
Notes and knowledge
- Plain Markdown with YAML frontmatter and tags
- Observations and typed relations
- Wikilinks and memory:// URLs
- Note types, schemas, validate, infer, and drift
- Dated facts: ask what was true when
- Locked notes and authorship
For your AI tools
- Write, read, edit, move, and delete over MCP
- Build context by following links
- Recent activity across projects
- Unix-style tools: ls, cat, grep, find, tail
- A built-in manual agents can read
Search
- Hybrid semantic and full-text search
- Optional reranking
- Metadata filters on any field
- One search across every project
- Chinese, Japanese, and Korean
Getting data in
- Claude and ChatGPT history
- PDF, Word, PowerPoint, and CSV
- Upload, drag and drop, and project ZIP
- MCP memory server JSON
Plugins and CLI
- Claude Code, Codex, OpenClaw, Hermes, and Pi plugins
- About 15 ready-made skills
- A CLI for every tool
- Session hooks any harness can call
Web app
- Markdown editor: tables, tasks, Mermaid, math
- Wiki and 39 live note components
- Comments and suggestions stored as CriticMarkup
- A merge view for you-and-an-agent edits
- Pocketbook inside Claude and ChatGPT
- WebMCP tools for browser agents
- Graph explorer, activity feed, offline-safe edits
Sharing and teams
- Public links with expiry and view counts
- Team workspaces for people and agents
- Five roles and invite-only projects
- Audit log
- Read-only and project-scoped API keys
Agents (beta)
- Instructions in a Markdown note
- Manual runs and watched folders
- Dashboard, run details, and a shared review queue
- Diffs, revert, restore, and undo a run
- Start and review runs over MCP
Start building with your AI.
Try it for 7 days. Your notes stay yours either way. Start building a knowledge base that grows with you.