# Nootio — full agent-readable description

> Nootio is **Governed Memory**: a controlled, auditable store of the rules, knowledge, and context that AI agents — and the people who work with them — are *allowed* to use. It is the memory layer of the Xalerate platform.

This document is written for AI agents and analysts. It describes what Nootio is, the full capability surface, and how an agent connects to and uses it. A shorter version is at /llms.txt.

---

## 1. The idea: Governed Memory

Most "AI memory" today is *ungoverned*: a model answers from whatever it absorbed in training, plus whatever happens to be in its context window. Nobody decided what it should remember, nobody can see why it believes something, and nobody can revoke a fact.

Nootio inverts that. Memory becomes something you **govern**:

- **Controlled** — you decide what goes in. Rules, decisions, lessons, preferences, facts, open questions, and the sources behind them.
- **Allowed** — agents read from a store you curated, not from an opaque blob. What an agent is *allowed* to use is an explicit, inspectable set.
- **Attributed** — every item has provenance: which source, which agent, who asked for it, when.
- **Tiered** — memory has retention semantics (durable vs working vs scratch), so transient operational state never pollutes long-term knowledge.
- **Auditable** — every create/delete an agent performs over the API is logged with the agent's identity.

That is "Governed Memory," and it is the working positioning of the product.

## 2. Platform role

Nootio is the memory layer of the Xalerate stack:

- **Nootio** — durable, governed knowledge: what people and agents have learned, decided, and agreed on.
- **BrainOrchestra (BO)** — the governed AI gateway in front of the models, enforcing data-residency and compliance routing.
- **Fleet Orchestrator** — coordinates fleets of agents.

Agents do their thinking through BO and coordinate through the Fleet Orchestrator; what persists — the knowledge and the agreements — lives in Nootio.

## 3. Usage modes

The same governed memory serves three modes from one store:

- **Personal** — an individual's knowledge base: capture anything, get summaries and synthesis, build durable memory.
- **Team** — shared spaces, shared billing pool, roles and permissions, team-visible knowledge and reports.
- **Agent** — AI agents connect over MCP with scoped identities and read/write the same memory the humans use.

## 4. How an agent connects (MCP)

- **Transport:** Streamable HTTP at `https://www.nootio.com/mcp` (stateless). Path-based labels (`/mcp/<label>`) allow multiple connectors.
- **Auth, two paths:**
  - **Named API keys** (`nootio_` prefix) for local MCP clients — created per agent in Settings → Privacy → MCP. The key's **label is the agent's identity**.
  - **OAuth 2.1** for remote connectors (PKCE/S256, dynamic client registration, resource indicators) — for hosted AI clients.
- **Scopes:** per key — `read` < `write` < `delete` < `admin`. Per-key space restrictions, per-key webhooks (standard or Slack format), soft-revocation with audit retention.
- **Identity tools:** `whoami` returns the key label, scope, and usage; `update_key_profile` lets an agent rename itself.

## 5. Capability surface

### 5.1 MCP tools — 104 tools, 11 resources

Categories: search; sessions (list/get/delete/update); reports (list/get/create/update/delete); memories & notes (create/get/update/delete, promote, correct, expire, set-expiry, bulk-create); references (CRUD, export); folders (CRUD, hierarchy, move); chains (CRUD, share, duplicate); spaces (CRUD, move, multi-space publish); handoffs (create/accept/complete/cancel/move/list/get); coordination teams (create/list/get/update/delete/set-members); concepts; favourites & read-state; trash (list_deleted, restore_item); audit (list_actions, confirm_action); profile & user-instructions; context assembly (get_context, get_active_context); import (CSV, Readwise); webhooks; analyze_text / analyze_url. Verify the live count: `grep -c '^  server.tool(' src/mcp/tools.js`.

11 `nootio://` resources expose sessions, reports, notes, memories, folders, and chains as readable URIs.

### 5.2 Agent identities & scoped keys

Each agent has its own named key (`mcp_keys` table). Per-key: scope, allow-delete/create/analyze flags, space restrictions (`mcp_key_spaces`), webhook config, last-used and words-used stats. Authorship label (`created_via_label`) is recorded on notes, sessions, and reports the agent creates. A user-level MCP toggle disables all of a user's keys at once and purges their embeddings.

### 5.3 Agent-to-agent handoff & coordination bus

Structured handoffs pass work context between agents with lifecycle tracking:

- **Fields:** objective, current_state, findings, open_questions, constraints, recommended_actions, result, typed result_references.
- **Lifecycle:** pending → accepted → completed (or cancelled). Identity-gated: only the accepting agent completes; only originator/accepter cancels.
- **Routing:** specific agent, any-of-team (first to accept wins), all-of-team broadcast (one child per member), or open.
- **Chaining:** `parent_handoff_id` links handoffs into workflow chains; `get_handoff(include_chain=true)` walks the chain.
- **Originator attribution:** an optional `originator` field records *who asked* for the work, distinct from `from_agent` (*who filed* it). Chain-children inherit the parent's originator. This is the difference between "an autopilot filed this" and "the human/agent who actually wanted it."
- **Goal-oriented tagging:** optional goal_id / workstream_id / contribution_type (advance, unblock, decide, verify, report, coordinate, watch, defer) / expected_outcome / blocked_by.
- **Coordination teams:** named groups of agent labels for any-of / all-of routing.
- A typed CLI (`@nootio/handoff-client`) and a Library workspace surface handoffs for humans.

### 5.4 Memory tiers

Every memory carries a `memory_class`:

- **durable** (default) — gardened (deduplicated, consolidated) and decay-scored over time.
- **operational** — working memory: excluded from gardening and decay; never deduped or depromoted. For agent heartbeats, status beacons, resume/working state.
- **ephemeral** — scratch: no decay, eligible for TTL auto-cleanup.

Spaces have a `default_memory_class`, so a status/coordination space can default everything to operational.

### 5.5 Memory lifecycle

Design principle: **"Episodes are truth. Memory is interpretation."** Sessions are raw episodes that never change; memories are interpretations with a lifecycle:

- **Status:** candidate / active / expired / corrected.
- Promotion (manual / auto on repeated reference / agent), TTL expiry, reactivation, bidirectional correction links, reference counting. Expired memories are never deleted — searchable and reactivatable.
- Typed memories: decision, lesson, question, preference, fact, task.

### 5.6 Credence / source stance

Sources carry a stance orthogonal to memory salience: `endorsed`, `neutral`, `counterweight`, `disputed`, `refuted` (default neutral). A **counterweight** is kept but never feeds report evidence — it surfaces as the red-team opposing view. `refuted` is excluded from synthesis but kept for provenance. Report generation splits sources by stance.

### 5.7 Spaces & permissions

Content lives in exactly one home space. Access is per-user and per-team with three roles (owner / editor / viewer). Publishing is *additive* — an item can be visible in multiple spaces, read-only in targets. Per-key MCP visibility gates (`access_mcp`) control which spaces an agent can see.

### 5.8 Audit trail

All create/delete operations via MCP are recorded in `mcp_audit_log` with the acting key's label (90-day retention). The `list_actions` tool queries it; webhooks can notify on key events.

### 5.9 Knowledge surfaces

- **Sessions** — analysed sources (summaries, conclusions, key takeaways, categorised references across 40+ categories).
- **Reports** — synthesis-first documents with inline `[[Reference]]` mentions, themes, visual diagrams, team sharing; superpower mode (multiple models synthesised); prompt-only and briefing modes.
- **Chains / Collections** — curated, slide-like decks of notes and references.
- **Notes & Memories** — annotations on sources and standalone curated memories.
- **Knowledge Graph** — force-directed view of sources, reports, references, and memories.
- **Daily Digest** & **Memory Gardener** — auto-summaries and a 4-pass engine that consolidates duplicates, audits freshness, and ages knowledge.

### 5.10 EU data residency

Per-user `data_region` tiers — Auto, USA, EU Cloud (Bedrock Frankfurt + EU models), EU Strict (Mistral France only), EU Sweden (Bedrock Stockholm). LLM, transcription (Deepgram/Voxtral), and embeddings route through EU infrastructure via the BrainOrchestra gateway. Consent is captured and versioned per region.

## 6. Ingest & synthesis

- **Capture:** paste text, file upload (PDF incl. per-page vision for charts, DOCX, PPTX, XLSX/CSV, images via vision extraction, audio, 60+ formats), URL analysis, email-to-analyse (forward to a unique address), live voice transcription, Apple Podcasts, and a Chrome web-clipper extension.
- **Process:** three-tier model pipeline (fast capture, mid-tier quick checks, full-tier deep extraction). Extracts summaries/conclusions/takeaways and categorised references; synthesises multi-source reports; answers questions over your own governed memory ("Noot" assistant).

## 7. Tech & operator

- Backend: Node.js + Express, PostgreSQL (+ pgvector), Redis (BullMQ). Frontend: React + Vite. Mobile: Expo / React Native. Deployed on Railway.
- AI providers: Anthropic Claude, Mistral, OpenAI — routed through the BrainOrchestra governed gateway with per-region compliance.
- **Operator:** Xalerate AB, org. nr 559575-8698, Karlstad, Sweden (a subsidiary of Helltech AB). Contact: niklas@nootio.com.

## 8. Links

- https://nootio.com/ — product overview
- https://nootio.com/how-it-works — how it works
- https://nootio.com/enterprise — enterprise
- https://nootio.com/information-flow.html — information flow & MCP tool inventory
- https://nootio.com/llms.txt — short version of this document
- https://nootio.com/privacy — privacy policy
- https://nootio.com/terms — terms of service