AMFS MCP
AMFS (SenseLab)
The SenseLab MCP server records an agent's decisions, scores them against task outcomes and tells the next run what worked
- 92 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The install script edits client configurations and the server stores decision traces and material the agent read in a local folder or a shared database.
Install
- Works in
Manual install
curl -sSL https://raw.githubusercontent.com/raia-live/amfs/main/install-mcp.sh | bashDetects installed clients and adds the `senselab` server. Read the script before running.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The install script edits client configurations and the server stores decision traces and material the agent read in a local folder or a shared database.
This is third-party code. Review the repository files before installing.
What it does
AMFS (SenseLab) is a memory and self-improvement engine for agents. Every run leaves a decision trace: what the agent was asked, what it read, what it decided, what it did and how it ended. The outcome moves confidence on everything the run relied on, and the next run of the same task gets a briefing and an action recommendation built from that data. The `senselab` MCP server installs with a script that finds installed clients (Cursor, Claude Code, VS Code, Codex, Windsurf, Gemini CLI, Claude Desktop) and adds an entry to each, while data lives in a `.amfs/` folder on your machine. The agent gets 40 tools, including `amfs_briefing`, `amfs_retrieve`, `amfs_write` and `amfs_commit_outcome`. `record_attempt` separates a failed attempt from a successful one so a stale entry does not get credit for the recovery. Every knowledge entry carries an evidence status: untested, validated, contested or discredited. There is the `amfs` package for Python, `@senselab-ai/amfs` for TypeScript and a Docker Compose setup with Postgres and pgvector for a shared server.
Who it is for. Teams running agents on recurring tasks such as deploys or incident triage who want the agent to stop repeating past mistakes.
Good fit when
- A deploy agent keeps trying a fix that already failed and such failures need to be stored
- Before working on a service an agent needs a briefing: what is validated here, what is discredited and what was tried
- A run is verified by CI or a human and that confirmation should weigh more than the agent's own word
- Several agents need to share experience through a common server with Postgres
- You want memory that forgets: beliefs fade without confirmation while experience stays
Not a fit when
- You need plain note memory with no outcome scoring: a simpler memory MCP server is enough
- Tasks are one-off and do not repeat: accumulated experience will not pay off
- Running an install script through `curl` is not acceptable: use the manual setup from the MCP guide instead
Example request
Before changing the checkout service request a briefing with amfs_briefing, and after the work record the outcome with amfs_commit_outcome as success.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
AMFS_HTTP_URL- Address of your own shared AMFS server, for example http://localhost:8080. Without it data is stored locally.
Limitations
The full loop with compiled briefings and action priors needs the server and Postgres, while the local filesystem mode gives only part of it. The install script runs through `curl | bash` and edits the configs of detected clients, so read it before running. For the agent to call briefing, write and commit at the right moments, add the instructions from `CLAUDE.md` to the project. The effect appears after several runs of the same task and an empty store gives recommendations nothing to rest on. The author calls it recursive only in the sense of reversible corrections and not open-ended capability gain.
How to disable. Run the install script with `--uninstall`: it removes the `senselab` entry from client configurations. Data stays in the `.amfs/` folder, which you delete by hand.
FAQ
How do I connect AMFS to Claude Code?
The `install-mcp.sh` script detects installed clients and adds the `senselab` server to each. Pick one with `--client`, for example `--client cursor`. Restart the editor afterwards.
Where is data stored?
In local mode in a `.amfs/` folder on your machine. For a shared server there is Docker Compose with Postgres and pgvector and the address is set with the `AMFS_HTTP_URL` variable.
How does the agent know what to do?
The `amfs_briefing` tool returns a briefing for an entity, and `retrieve` with candidate actions returns a recommendation: act, explore another option or escalate when every option already failed.
How does the validated status differ from untested?
They are different levels of evidence. An untested entry at confidence 0.9 and a validated one at 0.9 weigh differently in a decision and the agent can tell them apart.
Is there an SDK without MCP?
Yes. `pip install amfs` gives a Python library with `AgentMemory` and `@senselab-ai/amfs` covers TypeScript.
Discussion
No questions or reviews yet.
Related
MCP serversA local MCP memory server for coding agents: decisions, errors and facts stay on your machine, with semantic and graph recall
A long-term memory MCP server for agents: search, store and correct facts across sessions in the cloud or a local core
An MCP server and API of shared memory for several agents: facts, decisions and events live in one database with semantic search.
Local agent memory built on a corrections ledger: tell Claude Code once and a new session reads the rule before it starts work
ai-memory retains session knowledge in a Markdown wiki, exposes MCP retrieval and passes context between agents and machines.
An MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| AMFS (this card) | no | not required | 92 | |
| agentmemory | yes | not required | 30,340 | stores and retrieves agent memories; AMFS adds outcome scoring and an action recommendation |
| Basic Memory | no | not required | 4,566 | note memory on Markdown; AMFS builds trust in knowledge from run results |
README in short
The README explains why to close the agent learning loop, how the MCP server and Python library install, how to run a shared server and what the agent gets: decision traces, a briefing before acting and honest confidence scoring.