SMFS (Supermemory Filesystem) is an open-source layer that exposes a Supermemory container as a directory agents can treat like normal files. Agents use familiar tools (ls, cat, grep) instead of wiring a vector-database SDK or reloading entire chat histories into context. It is one variant of the broader Agent Memory Filesystems pattern and targets the same problem as the LLM Wiki pattern—long-term, stateful memory for agents—but optimizes for cloud-backed semantic retrieval and multi-modal ingestion rather than a local, RDF-validated markdown wiki.

Official docs: SMFS overview. Source: github.com/supermemoryai/smfs.

Design philosophy

SMFS is a RAG-augmented context layer: a knowledge graph and vector index sit behind a POSIX-like surface. The goal is to combine structured file-tree navigation with semantic ranking so agents spend fewer tokens and miss fewer relevant facts at scale. Background sync and graph maintenance are largely automatic (“dynamic dreaming” and entity updates on the Supermemory side), so the agent is not solely responsible for rewriting flat memory files after every session.

Architecture

Typical workflow:

curl -fsSL https://smfs.ai/install | bash
smfs login
smfs mount agent_memory

Files under configured memory paths (for example user.md, memory.md, or custom --memory-paths) are distilled and indexed by Supermemory; ordinary paths behave more like a synced file tree.

How agents experience it

• Semantic grep — Intercepted grep runs meaning-ranked search across the container; standard flags can force literal string match.
• profile.md — Virtual root file: a live, token-lean digest of the container the agent can cat instead of walking every subtree.
• Multi-modal — PDFs, images, and other types can be ingested; text representations become searchable through the same shell tools.

Comparison with Wiki CLI and Letta MemFS

Choose SMFS when you want semantic search hidden behind bash, multi-platform sync, integrations, and managed scale without maintaining a wiki toolchain. Choose the Wiki CLI when you want explicit, inspectable Declarative Knowledge, Procedural Knowledge via wiki check / wiki render, and a compounding markdown codebase under your control.

For the full cross-tool comparison, see Agent Memory Filesystems.

Sources