Ramet

Ramet: Workspace management system that gives AI agents isolated, forkable, and persistent filesystem environments.

Ramet is a workspace management system that provides persistent, forkable workspaces for AI agents. It allocates isolated writable workspaces derived from a shared immutable base, allowing frameworks to run parallel attempts and reset environments cleanly. Developers use the Python SDK, Go client, CLI, or HTTP interface to take snapshots, assign persistence policies to paths, and promote chosen states to disk, PostgreSQL, or S3-compatible storage. It is built for teams running repeatable evaluations, reinforcement learning rollouts, and multi-agent systems.

Ramet at a glance

Best for: Software engineer
Runs on: Linux, macOS, CLI, API, self-hosted, open source
Pricing: Free
Released: October 2026
Category: AI agent tool
Website: github.com/shenli/ramet

Key features

Fork isolated writable workspaces from a shared immutable base
Persist workspace files and local Git state across sessions and sandbox replacements
Define path persistence policies to separate durable, rebuildable, and ephemeral files
Reset filesystems cleanly with contamination verification between runs
Reuse prepared repositories, dependencies, caches, and warm pools
Mount workspaces to same-host Docker runtimes or custom sandboxes
Hydrate working-tree files lazily on first access with optional Linux FUSE backend
Store snapshots locally or publish them to PostgreSQL and S3-compatible storage

Use it for

Retaining agent code edits and Git commits across sandbox restarts
Running best-of-N coding agents with parallel isolated workspaces
Resetting reinforcement learning environments between rollouts in SkyRL
Reusing prepared dependency caches and repositories during agent evaluations

Alternatives to Ramet

Daytona — It provisions sandboxes in under 90 milliseconds with full security isolation.
LayerFS — Mounts complete workspaces and commits tool calls instantly.
Tigerless Agent Memory — Maintains agent history in portable, greppable Markdown files without API keys.
Agensh — Scales to 1,024 self-organizing agents without a central orchestrator.
Agent Executor — Declarative runtime for scaling stateful agent workloads with isolation.
OpenViking — Unifies agent memory, knowledge RAG, and skills into a single directory structure.
Starnet — Visual agent management with transparent, persistent on-disk logs and state.
MyAgent — It features built-in permission controls, long-term memory, and skill self-evolution.
AgentFS — Provides isolated filesystem access and instant snapshots via SQLite.
AgentDock — It coordinates multi-device workflows and provides secure environment access.
ReMe — It enables local-first, file-native long-term memory sharing across multiple AI agents.
Runspace — Executes builds, tests, and agent tasks on a cloud VM to keep your host machine clean.

All Ramet alternatives