New Introducing SMFS: Memory as a Filesystem →

Give your agent a filesystem. It gets a supermemory.

smfs mounts your Supermemory container as a real directory. Read it with ls, cat, and grep - sync, search, and a live profile included.

Macintosh Classic
FinderFileEditViewGoWindowHelp10:24
Macintosh
Sandbox
memory
Semantic search
Inside a mount, grep is transparently upgraded to semantic search. Run grep "auth flow" and get meaning-based results. grep -F falls through to standard grep.
Live profile
cat profile.md returns a live digest - synthesized on read from every memory in the container. Not stored, regenerated on demand.
Real filesystem
Real NFS on macOS, FUSE on Linux. No kext, no macFUSE. Pure Rust, #![forbid(unsafe_code)]. MIT or Apache-2.0.
// The Case

Why smfs exists?

Agents are inherently fluent in bash. They are built to navigate file structures, execute scripts, and pipe outputs. Yet, when we give them "memory", we force them to learn complex vector databases, manage API keys, and orchestrate external SDKs.

smfs collapses four moving parts (a vector DB, a memory service, a profile store, and an SDK) into a single mount point. We translate complex semantic operations into standard POSIX syscalls.

The filesystem is already the universal API for agents. We gave it memory, search, and a profile.

- SUPERMEMORY LABS

Your agent runs unix commands

Mount, search, and sync - the full memory lifecycle from your terminal.

agent@host:~/workspace$ cd ./memory
agent@host:~/workspace/memory$ grep "auth middleware issues" .
 
[smfs: executing semantic search across container]
./docs/architecture.md: We decided to use JWT for the auth middleware, but we've seen race conditions...
./slack_export/channel_eng.txt: @sarah mentioned the middleware is dropping connections on high load.
 
agent@host:~/workspace/memory$ _
// Use cases

Just files. Build anything.

smfs is a real POSIX mount, not another API. Every tool that already reads files works with it — your shell, your editor, your scripts, your agents. No SDKs to learn.

grep "GDPR clause" ~/smfs/legal/

Legal document search

Let agents search across contracts, compliance docs, and policies by meaning. Find the right clause without knowing the exact wording.

cat ~/smfs/finance/q4-reports/

Financial analysis

Agents read quarterly reports, invoices, and forecasts from a single mount. Ask questions across hundreds of documents.

grep "side effects" ~/smfs/research/

Research & documentation

Point agents at your research library. Semantic search across papers, internal wikis, and technical docs in any format.

ls ~/smfs/support/tickets/

Support & knowledge bases

Support agents read from a shared knowledge base that stays in sync. Every agent sees the same docs, always up to date.

cp onboarding.pdf ~/smfs/hr/

HR & operations

Drop employee handbooks, org charts, and policy docs into the mount. Agents index and search them instantly.

cat ~/smfs/profile.md

Personalized agents

Every agent reads a live profile synthesized from all memories. Context that follows your users across sessions.
// Benchmarks

Agents retrieve more with less

220 audited xAFS questions across Claude Code and Codex. Same agent, same documents — just add smfs.

0.0%
tokens saved
Codex 84M → 48M · Claude 72M → 24M
0/220
correctness
Claude 100/110 · Codex 101/110 (SMFS)
0
tokens per correct answer
-53% vs filesystem (766k)
0.0%
tokens saved with Codex
Without smfs
84,289,041
With smfs
48,059,042
Pass rate (SMFS)
Claude 100/110 · Codex 101/110
AgentTool callsTotal tokensCorrectnessTokens / correct
Claude862 → 345 -60%72M → 24M -66%99 → 100/110 +1%727k → 241k -67%
Codex1377 → 1137 -17%84M → 48M -43%105 → 101/110 -4%803k → 476k -41%
// Any format

Drop the file. Then grep across all of them.

PDFs, videos, screenshots, audio, docs - drop the raw files straight into the mount. No OCR. No transcription. No PDF parser. No chunking. The pipeline you'd have built is gone, and grep just works across every format.

~/desktop
$ _
~/smfs/0 / 5 indexed
  • 01No OCR pipeline.
  • 02No transcription step.
  • 03No PDF parser.
  • 04No chunking job.

Mount it.

001

smfs mounts supermemory as a real folder. Drop files in. Point your tools at it.

Semantic by default.

002

grep returns meanings, not substrings. Ask for "my pdf" - get the invoice you saved.

Cloud-synced. Agent-ready.

003

Files live in the supermemory cloud. Every agent sees the same folder, in sync.
[VIEW: SCHEMATIC_03]
SYS.MEM.ALLOC // OK
LAYER 01 // SURFACE
Any tool
The surface your agent already knows.
[cat][grep][ls][vim][VS Code][git][tar][find]
↓ POSIX SYSCALLS
LAYER 02 // SYSTEM
OS kernel
NFSv3 client on macOS / FUSE on Linux.
No kernel extension. No macFUSE. No security prompts.
  • macOS backend01
    NFSv3 over localhost
    Pure-Rust NFS server on 127.0.0.1:<auto>, mounted via /sbin/mount_nfs.
  • Linux backend02
    Kernel FUSE
    Zero-setup via the fuser crate. NFS available as alternate backend.
  • POSIX03
    Full inode semantics
    symlinks · hard links · rename · chmod · chown · truncate · statfs
↓ RPC / FUSE OPS
LAYER 03 // MIDDLEWARE
smfs daemon
Pure Rust, one static binary.
  • Language01
    Pure Rust
    #![forbid(unsafe_code)] · MIT or Apache-2.0 · MSRV 1.80
  • Local cache02
    SQLite
    Survives daemon restarts · enables offline reads
  • Sync engine03
    Bidirectional
    Pull on cache miss / TTL · push dirty writes with exponential-backoff retry
  • Semantic grep04
    Transparent wrapper
    Installed on first mount · falls through to standard grep outside mounts · grep -F always literal
↓ HTTPS
LAYER 04 // CLOUD
Supermemory
Your container · embeddings, ranking, memory.
  • Virtual file01
    profile.md
    Live-generated AI memory profile at the mount root · read from /v4/profile
  • Isolation02
    One mount per container tag
    Clean separation of data scopes · credentials per project or global

mountmemory.

One binary. No SDK. No vector DB to babysit.

READYinstall.sh
$ curl -fsSL smfs.ai/install | sh
$ smfs mount my-project
# → grep away. we'll do the vectors.