MemFS (memory filesystem), also called a context repository in Letta marketing, is Letta Code’s git-backed memory system. An agent’s long-term state lives as a tree of markdown files in a local Git repository—human-readable, diffable, and versioned—instead of opaque memory blocks or a proprietary vector store alone. It is one variant of the broader Agent Memory Filesystems pattern and addresses the same agent-memory problem as the LLM Wiki pattern and Supermemory SMFS, but optimizes for deterministic, personalized coding agents with clear always-on vs on-demand context.

Official reference: MemFS | Letta Docs. Overview: Letta Code memory. Announcement: Context repositories blog.

Design philosophy

MemFS treats memory as a repo the agent edits with bash, then commits. Letta API agents clone to ~/.letta/agents/<id>/memory and push commits to sync with Letta Cloud; local-mode agents use ~/.letta/lc-local-backend/memfs/<id>/memory (or $LETTA_LOCAL_BACKEND_DIR). Progressive disclosure keeps filenames and YAML description frontmatter visible while full file bodies load only when needed—similar in spirit to agent “skills” with SKILL.md metadata.

MemFS is default for new Letta Code agents (v0.15+) on the Letta API and in local mode. It is not available on Docker-server deployments (legacy memory blocks remain). Migrate with /memfs enable.

The system/ directory

The distinguishing layout rule:

• system/ — Every markdown file here is loaded in full into the system prompt on every turn (persona, user preferences, core project facts).
• Outside system/ — The agent sees the tree and descriptions; contents are pulled in only when relevant.

The agent is expected to move files between system/ and the rest of the tree as it learns what must stay hot vs what can stay lazy. Changes take effect after a Git commit (and push for API agents).

Example frontmatter on a memory file:

---
description: Coding preferences learned during /init
---

Bootstrap and evolution

• /init — Subagent workflow that crawls the codebase and captures preferences; seeds the repository.
• Agent-driven hierarchy — Unlike SMFS’s heavier background graph updates, MemFS relies on the agent (and you, in an editor) to curate structure over time.
• Human-in-the-loop — Open the memory folder in VS Code or Cursor and edit markdown directly; Git history provides audit and rollback.

Comparison with Wiki CLI and Supermemory SMFS

Choose MemFS when you run Letta Code and want Git-auditable, human-editable agent memory with a crisp hot/cold split and an always-on system/ slice. Choose the Wiki CLI when you want a shared, schema-checked wiki with Second Brain-style linking and semantic web tooling independent of a single agent runtime.

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

Sources