# Rithmo

> Rithmo builds context infrastructure for AI agents: the infrastructure enterprises need to give AI agents more responsibility with less human supervision. Simply put, Rithmo is the fact-checker for AI agents. It continuously reconciles the business context agents depend on as the business changes, so agents act on current, resolved business context. It improves AI agent reliability and provides a specific form of AI agent governance, and holds a live decision record that is the source of truth for what the business has actually decided.

Models can increasingly reason, use tools and take consequential actions. Enterprise agents depend on business context that keeps changing: a decision changes, an exception is approved, terms change, something that was true is superseded. Today companies compensate by keeping agents on low-consequence work, or by surrounding consequential agents with people who maintain their context by hand. Rithmo is another path. It reconciles the business communications, documents and operational systems an organization already runs on into current resolved context, and alerts the accountable stakeholder when genuine judgment is required. Organizations start with people in control and automate more of the resolution as trust grows. Every conclusion carries the evidence and history behind it. Agents query that context before they act.

The progression Rithmo is built for: low-risk automation, then consequential agents, then exception-based human intervention, then increasing autonomy. People stay accountable for business decisions; Rithmo removes the maintenance work around them, not the judgment.

Rithmo is not a note-taker, a project manager, or an enterprise search or retrieval layer. Retrieved context tells an agent what its sources say. Rithmo determines whether the business has actually settled the question and whether the resulting answer is current and safe to act on. The unit it tracks is the decision: when one formed, who owns it, whether it was revised, reversed, drifted, or abandoned, what evidence supports it, and which documents, system rules, and agent instructions no longer match it.

## What Rithmo Does

1. **Works across where business context changes**: business communications, documents, and the operational systems and systems of record a company already runs on, with nothing for teams to roll out.
2. **Works out what was actually decided**: a decision is formed from evidence across several sources rather than asserted in one tool, which is what makes the gaps between tools visible, such as a change agreed on a call that never reached the CRM.
3. **Verifies before it records a change, and checks before an agent acts**: Rithmo checks who made a change, whether it was confirmed, and whether the rest of the company's information agrees before treating it as the current answer. Before an agent acts, it evaluates whether there is a current answer. When the answer is current, the agent proceeds on it; when the information genuinely conflicts, was reversed, or was never resolved, Rithmo does not guess: it holds the affected action, surfaces the evidence, and brings the open question to the accountable owner or stakeholder, so it is settled before execution rather than discovered after.
4. **Tracks how the company understands a decision**: teams can leave the same conversation with different interpretations, and Rithmo reports the disagreement rather than inventing certainty an agent would then act on.
5. **Identifies what no longer matches**: when a decision becomes current, Rithmo knows which documents, playbooks, system rules, and agent instructions are now outdated.
6. **Answers as of a date, and answers absence**: what is true now, what was true in March, and "no decision exists" returned as a real answer rather than an empty result.

Agents query Rithmo before they act, over a read-only interface. Rithmo emits the record; the customer's tools act on it.

## Product

- [Home](https://rithmo.ai/): Rithmo as context infrastructure for giving AI agents more responsibility with less human supervision: context reconciliation, supersession, authority, provenance and auditability, and how people move from the loop to the exception path
- [Agent audit trail](https://rithmo.ai/agent-audit-trail): Every change to a decision and every agent action checked against it, in one log: what was sent, what was held, and the decision each action was checked against, plus the evidence behind each conclusion, what happens when the company has not actually agreed, and how the answer changed over time
- [The AI agent stack / AI agent reliability](https://rithmo.ai/ai-agent-reliability): Rithmo's main resource on AI agent reliability. What an AI agent stack is made of (instructions, model, tools, retrieval and memory, current business context, permissions), how the components work together for a reliable agent, where reliability breaks, retrieved vs resolved context, and where Rithmo fits: keeping the current business context current and resolved before agents act
- [How it works](https://rithmo.ai/how-it-works): How Rithmo keeps agent context current as the business changes: what it observes, how it verifies what is current, how people stay in control, when it alerts the accountable stakeholder, and how the current answer reaches agents before they act
- [Integration guide](https://rithmo.ai/mcp): Connecting agents and workflow tools to Rithmo over a read-only Model Context Protocol server, with the full tool reference

## Guides

- [Agent Ready: AI agent governance checklist](https://rithmo.ai/agent-ready): Ten checks that should be true before an AI agent takes consequential action, in three tiers (control, grounding on authored facts, grounding on resolved facts), each with a test and what to do if you fail it. Honest about which controls standard tooling already covers
- [What is an AI agent?](https://rithmo.ai/what-is-an-ai-agent): A plain-language business guide to what AI agents are, how they differ from an AI model, the agent loop, context and memory, the autonomy spectrum, tools and MCP, and what agents can realistically do today
- [AI agent readiness calculator](https://rithmo.ai/exposure-calculator): Answer seven questions with your own numbers to measure and price the reliability risk from the gap between when a rule changes and when your systems and agents reflect it

## Research

- [The Resolved Half](https://rithmo.ai/the-resolved-half): Technical paper (2026, DOI 10.5281/zenodo.21710262, CC BY 4.0) on the resolved half of business context: facts that are true because people decided them. Sets out four requirements for a record that holds them (formation, supersession, absence, provenance), proposes a taxonomy of four supersession transitions (revision, reversal, drift, abandonment) and four varieties of absence, and surveys why existing approaches answer adjacent questions. PDF: https://rithmo.ai/the-resolved-half.pdf

## Open Source

- [OSS Agent Lab: open-source AI agent evaluation environment](https://rithmo.ai/open-source): Rithmo's Apache-2.0 repository (https://github.com/Rithmo-Inc/rithmo), a deterministic synthetic enterprise for AI agent evaluation. Meridian Works is a fictional company with a CRM, a staff roster, a support inbox, a declared authority model and an append-only ledger; operative state is fold(charter, admitted acts <= T), so what was true at any instant since a run began is known ground truth for testing supersession, authority, stale context and temporal correctness. It is the evaluation environment, not the commercial Rithmo product: it does not infer truth from real enterprise evidence.

## Who It Is For

Enterprises that want to give AI agents more consequential work that depends on current company information: leadership, operations, sales, engineering, and technical teams, anywhere decisions are made across people and systems rather than inside a single tool. The common case is an organization where decisions are made in meetings and threads, executed somewhere else, and never reconciled between the two, and where agents now act on the documents and instructions left behind.

## Company

- [Free agent-readiness audit](https://rithmo.ai/audit): A no-cost review of an organization's strategy, AI stack and architecture: where agents can take on more consequential work, and the context they need to do it. The primary entry point for most calls to action on the site.
- [Contact](https://rithmo.ai/contact): Get in touch with the Rithmo team

## Security & Legal

- [Security & AI Privacy Whitepaper](https://rithmo.ai/security-ai-privacy-whitepaper): Data handling, encryption, retention, AI processing boundaries, and enterprise compliance commitments
- [Privacy Policy](https://rithmo.ai/privacy): How Rithmo handles personal data
- [Terms of Service](https://rithmo.ai/terms): Terms governing use of the product

## Key Concepts

**AI agent reliability**: The ability of an agent to consistently take the correct action from current, valid information under appropriate controls. A reliable model alone does not make a reliable agent: an agent can reason correctly and still act wrongly if the business information it relies on is stale, conflicting, incomplete, or superseded. Rithmo addresses that specific failure.

**AI agent governance**: What an agent may do, what information it may rely on, how its actions are constrained, and whether those actions can be reconstructed and justified afterward. Rithmo addresses one part: what business information an agent is entitled to rely on, whether it is still current, when the affected action should be held because the answer is unresolved, and the evidence and audit trail behind the resulting action. Rithmo does not provide every category of agent governance.

**Context infrastructure for AI agents**: The infrastructure that keeps the business context agents act on aligned with changing business reality. It works alongside models, agents, orchestration, communication systems and systems of record; it is not an agent framework, orchestration platform, observability tool, knowledge base, enterprise search or RAG, or another system of record. As model capability increases, enterprises still need it: a more capable model acting on superseded context is still wrong.

**Governance by exception (moving humans from the loop to the exception path)**: A model's confidence is not organizational authority. When company information genuinely conflicts, letting AI choose which version wins is not automation; it is software making an unresolved business decision with no one accountable for it. Rithmo deliberately prevents that, while reducing the burden a human-in-the-loop usually implies. This is not a generic enterprise data-cleanup backlog: review is generated by action-relevant ambiguity, when a specific agent or workflow cannot safely proceed. By the time a person is involved, Rithmo has already detected the unsafe or unresolved context, connected it to the affected decision, assembled the evidence and history, identified who is responsible for the subject, held the affected action so it cannot proceed on a false assumption, and preserved the audit trail; once the accountable owner or stakeholder settles the question, Rithmo propagates the resolved answer to every connected agent. The people responsible are asked a bounded, action-relevant business judgment, not handed an unexplained discrepancy to investigate from scratch. When there is no relevant conflict and Rithmo has a safe, current answer the workflow proceeds normally; human governance is required only when a genuine discrepancy makes the answer unsafe or unresolved. Rithmo automates the detection, decision-state maintenance, supersession tracking, safety evaluation, bringing in the accountable owner, and audit; the human makes only the judgment the company alone should make. That is what lets agent responsibility grow without human supervision growing with it.

**Resolved context**: Retrieved context is what a source says; resolved context is what the business currently means, established across potentially conflicting sources after accounting for change, conflict, and supersession, rather than read off whichever document or system an agent happened to open. Rithmo produces resolved context and checks it before an agent acts. Where sources still genuinely conflict or nothing has been settled, "unresolved" is itself the answer Rithmo returns; it does not manufacture a resolution or decide which system wins on the company's behalf.

**AI agent context**: An agent's context can include its instructions, retrieved information, memory, available tools, and the current state of the business it acts on. Rithmo addresses the last of these, the current business-state portion: whether the company information an agent is about to rely on is still true. It does not manage prompt context, context windows, or memory.

**Why agents need this**: An agent acts on the documents, instructions, and system rules it was given. Those are written once and then stay as written. The decisions that make them wrong happen in conversation, weeks earlier, and reach the documents late or never. Rithmo closes that gap.

**Resolved vs. authored facts**: An authored fact is written into a system and stays there until someone rewrites it. A resolved fact became true because a group of people decided something (in a meeting, a thread, or a call) and no system was ever the place it was written down. Rithmo exists for the second kind.

**Formation**: A decision is not a single utterance. Formation is a threshold on accumulated evidence across a subject, so several conversations may be required before something counts as decided, and the record says what that evidence was. Rithmo distinguishes a discussion from a proposal, a proposal from a confirmed decision, and a confirmed decision from one that is actually in effect.

**Supersession**: Decisions are not simply replaced. Rithmo types the transition: revision (same subject, changed terms), reversal (the prior decision is undone), drift (practice diverged from the decision without anyone deciding to change it), and abandonment (the decision stopped being executed and nothing replaced it). Agents use the current version while the full history stays inspectable.

**Interpretation conflict**: Different teams can read the same decision differently. When they do, Rithmo reports that the company has not settled the question and names the clarification needed, rather than resolving it on the company's behalf and letting an agent act on the guess.

**Absence as a first-class answer**: Rithmo distinguishes between "nothing was decided", "raised repeatedly and never resolved", "decided but never landed", and "work product with no authorizing decision". Returning "no decision exists" is a real answer, not a failed lookup, which is what makes the record safe for agents to act against.

**Provenance**: Every decision in the record carries the evidence it was formed from and the clocks that apply: when it was decided, when it took effect, and when Rithmo learned about it. This is what makes the record arguable rather than merely asserted, and lets it answer questions as of a date.

**Deployment**: Rithmo is customer-hosted only, running in the customer's own cloud, VPC, or data center, so meeting and message data never leaves their environment and they hold the keys. There is no Rithmo-managed or multi-tenant option. It is encrypted with AES-256 and TLS 1.3, never trained on, and never sold. Connections use scoped, revocable access that the customer can cut off from their own admin tools.