# SenseLab (AMFS)

> AMFS is an open-source Agent Memory File System that gives multi-agent AI systems shared, versioned, outcome-validated memory with Git-like branching and a cognitive layer for continual learning.

SenseLab builds AMFS (Agent Memory File System), an open-source cognitive layer for multi-agent AI systems, licensed under Apache 2.0. The core idea is that agents should not just store context — they should validate each other's findings, build on shared knowledge, and improve based on real production outcomes. The project is maintained by SenseLab Corp and backed by Speedrun, with a cloud-hosted Pro tier available at sense-lab.ai alongside the open-source engine.

## What It Is

AMFS is a shared memory infrastructure for AI agent fleets. Where vector databases match similarity, AMFS keeps versioned knowledge that carries confidence scores, provenance, and outcome feedback. When a deploy succeeds or an incident fires, AMFS records what the agent read, what it chose, and what happened — shifting confidence based on real results rather than synthetic runs. The result is a memory layer that compounds: knowledge gets sharper with every production outcome, and every decision trace automatically becomes SFT/DPO training data.

## How the Cognitive Layer Works

AMFS structures agent memory around three operations:

- **Discover** — Before acting, an agent queries AMFS for what the fleet already knows. Results come back ranked by confidence with full provenance: which agent wrote them, when, and how trustworthy they are.
- **Handoff** — A finding written by one agent is readable by every other agent within milliseconds, across frameworks, sessions, and machines.
- **Learn** — When outcomes are committed, confidence on related entries adjusts. The system knows what to trust and what to question.

Rooms provide shared spaces where teams and agents coordinate around a project or workflow. Agents can see what others are working on, what they decided, and why — enabling cognitive coordination before any decision runs.

## Architecture and Open-Source Model

AMFS is a monorepo with a layered architecture. The OSS edition includes the full memory engine: versioned writes, confidence scoring, outcome feedback, causal traces, knowledge graph, hybrid search (full-text + semantic + recency + confidence), git-like timeline, SDKs, storage adapters, HTTP API, MCP server, and CLI.

Key packages:
- **Python SDK** (`pip install amfs`) — core `AgentMemory` class
- **TypeScript SDK** (`npm install @senselab-ai/amfs`) — full async API
- **MCP Server** (`pip install amfs-mcp-server`) — first-class support for Cursor, Claude Desktop, and Claude Code
- **HTTP Server**, **Core Engine**, and **CLI** available as separate installable packages
- **Storage adapters** for filesystem (default), PostgreSQL, S3, and HTTP remote backends
- **Framework connectors** for CrewAI, LangGraph, LangChain, AutoGen, and AWS Strands Agents

The Pro cloud tier (AMFS Pro) adds branching, merge, pull requests, access control, tags, rollback, cherry-pick, fork, multi-tenant isolation, immutable decision traces, the intelligence layer (Cortex), and a web dashboard. The README describes the distinction as: "OSS = single-branch repo with full history. Pro = GitHub."

## Setup Path

Getting started requires three steps: sign up for an account (free tier, no credit card required), generate an API key, then connect via MCP or SDK. For MCP-compatible clients like Cursor or Claude Desktop, a single `curl` command installs the MCP server. No framework rewrites are needed — AMFS plugs into whatever stack is already running.

## Update: v0.2.0 — Continual Learning Infrastructure

The latest GitHub release is **v0.2.0**, published May 29, 2026, titled "Continual Learning Infrastructure." The repository was last pushed to on August 10, 2026, indicating active development. The project has 59 stars and 3 forks on GitHub. Blog posts from the SenseLab engineering team, published in June 2026, cover the cognitive layer concept, continual learning architecture, and cross-agent context portability — signaling active product direction toward making agent fleets collectively smarter over time.

## Features
- Shared agent memory across frameworks, sessions, and machines
- Confidence scoring weighted by real production outcomes
- Outcome-validated learning (SFT/DPO training data auto-generated from decision traces)
- Rooms for agent coordination and team knowledge sharing
- Hybrid search: full-text + semantic + recency + confidence
- Causal explainability via explain() — shows which memories drove a decision
- Git-like timeline with full audit trail of every read, write, and outcome
- Branching, diffs, pull requests, rollback, and named snapshots (Pro)
- Knowledge graph with auto-materialized relationships
- Access control per branch, user, team, or API key (Pro)
- MCP server for Cursor, Claude Desktop, and Claude Code
- Python SDK and TypeScript SDK
- Storage adapters: filesystem, PostgreSQL, S3, HTTP remote
- Framework connectors: CrewAI, LangGraph, LangChain, AutoGen, AWS Strands
- Docker support
- CLI tools
- Multi-tenant isolation (Pro)

## Integrations
Cursor, Claude Desktop, Claude Code, CrewAI, LangGraph, LangChain, AutoGen, AWS Strands Agents, OpenAI, n8n, PagerDuty, GitHub, Slack, Jira, MCP (Model Context Protocol)

## Platforms
MACOS, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Open Source, Free tier available

## Version
v0.2.0

## Links
- Website: https://www.sense-lab.ai
- Documentation: https://raia-live.github.io/amfs/
- Repository: https://github.com/raia-live/amfs
- EveryDev.ai: https://www.everydev.ai/tools/senselab-amfs