MCP Setup

curl -sSL https://raw.githubusercontent.com/raia-live/amfs/main/install-mcp.sh | bash -s -- --api-key <your-api-key>

{
  "mcpServers": {
    "senselab": {
      "command": "uvx",
      "args": ["--refresh", "amfs-mcp-server-pro"],
      "env": {
        "AMFS_HTTP_URL": "https://amfs-login.sense-lab.ai",
        "AMFS_API_KEY": "<your-api-key>"
      }
    }
  }
}

claude mcp add senselab \
  -e AMFS_HTTP_URL=https://amfs-login.sense-lab.ai \
  -e AMFS_API_KEY=<your-api-key> \
  -- uvx --refresh amfs-mcp-server-pro

amfs_set_identity("checkout-agent", "Fixing retry logic")

---
name: senselab-memory
description: "Use SenseLab as persistent memory. When the user asks what you know, remember, or have saved about anything — including personal preferences — call amfs_retrieve BEFORE answering. Save durable facts and decisions with amfs_write. Also covers the session lifecycle: set identity, briefing, commit outcome."
---

# SenseLab Skill

SenseLab is persistent memory shared across every tool, session, and machine on
this account. Anything saved from one client is recallable from another.

It stores **whatever was saved** — personal facts and preferences ("what food do
I like?", names, dates, how the user likes things done) just as much as code
decisions and runbooks. Never tell the user that SenseLab is only for work or
agent memory.

## Recall first (the rule that matters most)

When the user asks what you know / remember / have saved about **anything**,
call `amfs_retrieve` before you answer:

`amfs_retrieve(query="<the user's own words>")`

`entity_path` and `key` are optional — it searches by meaning across everything
you can see, so a plain phrase is enough. Never tell the user you have no memory
of something without running `amfs_retrieve` first.

`amfs_read` and `amfs_recall` need an **exact** key. A miss there means "try
`amfs_retrieve`", not "nothing is stored".

## Trigger conditions
- Starting any session where SenseLab MCP is connected
- User asks to remember, store, or retrieve something
- User shares a durable fact, preference, or decision
- After making a significant decision or completing meaningful work
- When context from a previous session may be relevant

## Session workflow (always follow this order)

1. **Set identity** — always first, every session. Use a stable kebab-case name that persists across conversations about the same domain. Pass `model=` so decision traces record which model made each call.
   `amfs_set_identity("your-agent-name", "what you're doing this session", model="your-model")`
   Good names: `api-agent`, `infra-agent`, `dashboard-agent`
   Bad names: `fix-button` (too specific), `agent-1` (meaningless)

2. **Check memory before working** — reuse what is already there instead of re-deriving it.
   `amfs_briefing(entity_path="repo/module")` for compiled context on a codebase.
   `amfs_retrieve(query="<topic>")` for anything you cannot address by path.

3. **Work** — read before decisions, write after meaningful ones

4. **Commit** — always at session end, it's free
   `amfs_commit_outcome("task-ref", "success|minor_failure|failure", task_input="<the request>", response_text="<your final answer>")`

   `task_input` is what the agent was asked, in the words it arrived in;
   `response_text` is what it answered, in full. SenseLab sees only its own
   tool calls, so neither is captured unless the agent passes it — and a judge
   grading the answer reads nothing without `response_text`.

## Entity path conventions
Use hierarchical `{project}/{module}` paths to scope knowledge. Avoid generic paths like `project` or `app` — they become unsearchable.

Examples:
- `myapp/auth` — authentication module
- `myapp/api` — API layer
- `myapp/checkout` — checkout service
- `user/context` — personal or cross-project context
- `user/preferences` — working style, communication preferences

## Key naming conventions
- `pattern-` reusable patterns
- `risk-` known bugs or blockers
- `decision-` architectural or strategic choices
- `task-summary-` what was done and why
- `action-` log of actions taken

## Cost model
- Read = 1 op (`amfs_read`, `amfs_search`, `amfs_list`, `amfs_briefing`)
- Write = 2 ops (`amfs_write`, `amfs_record_context`)
- Commit = free (never skip)

Prefer `amfs_briefing` over multiple individual reads — one briefing returns compiled, ranked knowledge for an entity.

## What to write
- Durable personal facts and preferences the user shares, and anything they say "remember…" about
- Decisions with rationale (why, not just what)
- Risks or blockers discovered
- Patterns worth reusing
- Task summaries after meaningful work
- User decisions that affect direction

## What to skip
- Trivial changes (typos, renames, comments)
- Anything recomputable from the codebase in seconds
- One-sentence observations — batch related ones into a single richer entry

## Memory types
- `fact` — stable, verified knowledge (default)
- `belief` — hypothesis or unverified observation, use confidence < 0.9
- `experience` — action log, decays slower, good for what was done and why

## Confidence guidelines
- `1.0` — verified, tested in production
- `0.7–0.9` — high confidence, not yet verified
- `0.4–0.6` — hypothesis, needs validation
- `< 0.4` — speculative, early signal

## Good to know
- `amfs_briefing` still answers on a new store: if the Memory Cortex has not compiled a digest yet, it falls back to recent raw entries for that entity
- `amfs_commit_batch` writes several entries as one atomic commit — use it when you have a group of related entries to save
- Rooms are shared workspaces with teammates' agents. `amfs_my_rooms()` lists them; entries on a room's entity are visible to every member, so check before writing anything private

Machine A (Cursor / Bruno):
  → Reviews checkout-service PR
  → amfs_write("myapp/checkout", "risk-race-condition", "...")
  → amfs_commit_outcome("PR-456", "minor_failure")

Machine B (Claude Code / Alice):
  → Starts working on checkout-service
  → amfs_briefing(entity_path="myapp/checkout")
  → Sees Bruno's risk signal with boosted confidence
  → Avoids the same issue

"args": ["amfs-mcp-server-pro@0.7.12"]