Before agentsfs
- Agent researches something useful.
- The session ends.
- Context disappears or stays trapped in one vendor memory.
Agent-native filesystem, owned by the user
A portable filesystem for compounding agent knowledge.
agentsfs gives agents a user-owned repo for context, sources, decisions, files, and next steps that survive individual sessions and projects, plus CLI + MCP tools for search, backlinks, doctor, prompts, and progressive disclosure.
Git does not write your commits. It makes committing structured and cheap. agentsfs does the same for compounding knowledge.
agentsfs/
|-- AGENTS.md self-describing root for any agent
|-- projects/
| |-- claim.md timeline, status, next useful context
| `-- research.md dense notes with sources and wikilinks
|-- entities/
| |-- [[Acme Co]].md current understanding and references
| `-- [[Policy]].md terms, citations, open questions
|-- files/
| `-- report.pdf described in this folder's INDEX.md
`-- .agentsfs/ derived indexes, rebuilt from files + description: "What this file is for" + sources: ["url", "email", "document"]
The gap
A session can build real judgment. The next one should inherit it.
Today, agent work dies in chats, vendor memories, scattered files, and half-written notes. In an hour, an agent can research a company, untangle a claim, or assemble a working model of a domain. Then the session ends, and the next one starts cold.
Vendor memories are locked to one harness. Plain folders are portable, but they do not tell an agent where to write, how to find context, or how to leave work another tool can pick up.
What it is
A boring, durable contract: files first.
An agentsfs instance is a plain git repo. It holds any file type - notes, PDFs, spreadsheets, images, code, datasets - while markdown descriptions, frontmatter, and wikilinks make the whole tree legible to agents and humans.
The payoff is simple: your memory survives model changes, tool changes, vendor changes, and time.
01 No LLM inside
Your own agent does the synthesis. agentsfs makes the work obvious, structured, and cheap.
02 Files are truth
Indexes, caches, and embeddings are derived. The substrate survives as ordinary files.
03 Git is the backbone
History, provenance, offline work, and remotes come from git, with clone as the exit ramp.
Concrete workflow
Research once. Let every future agent pick up from the same substrate.
Research a company once. Preserve the sources, people, claims, PDFs, decisions, and open questions. The next agent can orient from the tree, follow wikilinks, and continue from the same files.
After agentsfs
- Agent writes decisions, sources, files, and next steps into a portable repo.
- Git preserves history, provenance, offline work, and sync.
- Another agent opens the same substrate later and knows where to continue.
agentsfs init tree to orient write dense notes commit to git doctor guides cleanup
How it works
Start with plain files. Add CLI, MCP, search, backlinks, and doctor when you want speed.
tree Progressive disclosure: folders and files with one-line descriptions and freshness from git.
search Full-text and semantic retrieval over the substrate, exposed to agents through CLI and MCP.
backlinks Find every reference to a wikilinked entity, document, project, or file.
rename Refactor a knowledge file and rewrite its wikilinks in one deterministic pass.
doctor Flag missing descriptions, dead links, sparse stubs, or duplicate-looking files.
prompts Onboarding and gardening instructions teach any agent how to maintain the substrate.
mcp The same capabilities are exposed through an MCP server for compatible agent clients.
60-second orientation
The contract is just files and git; the toolkit makes the common agent moves cheap.
$ agentsfs init
$ agentsfs tree
$ agentsfs search "claim history"
$ agentsfs doctorEvery harness
Point any agent at the same substrate.
Claude Code Codex OpenClaw Any MCP client Local scripts
Cross-harness neutrality is the point: when a new tool, model, or agent surface appears, the user does not need to re-establish context. The agent can orient from the tree and work against the same files.
AI power users Dev teams using multiple coding agents Researchers Operators with long-running projects Agent framework and MCP users
You own it
Plain files, real git, any remote.
Use no remote, a self-hosted bare repo, GitHub, GitLab, or a future hosted service. The sync story is git push and pull. The exit story is git clone.
$ git clone your-agentsfs The whole substrate comes with you: markdown, media, datasets, history, and the conventions that tell the next agent what everything is for.
FAQ
The short version.
Why not vendor memory?
Vendor memory is harness-locked and opaque. agentsfs belongs to the user and works across harnesses.
Why not Obsidian or Notion?
Those tools are app-first. agentsfs is a neutral filesystem contract any harness can adopt.
Why git?
Git gives history, provenance, offline work, remotes, push/pull sync, and clone as the exit ramp.
Does this require hosted sync?
No. The core is plain files and git. Hosted services can compete on convenience, not captivity.
How does MCP fit in?
The CLI and MCP server expose the same reading, writing, and navigation capabilities to agents.
Why not only a vector database?
Search and embeddings are useful derived indexes. The source of truth stays ordinary files.
Hosted, when you want it
Get early access to hosted agentsfs.
Synced agent memory, hosted search, and a web viewer are the convenience layer. Open-source core first. Hosted instances next.
Loading hosted waitlist...