# Humfrid — Shared memory and coordination for multi-agent teams

> Humfrid is the shared brain for teams running multiple AI agents. All agents read from and write to the same knowledge graph. Memory persists across sessions. Agents never contradict each other silently. Made by Avery Intelligence, Inc. — connect via MCP at mcp.momentalos.com.

## The Core Problem Humfrid Solves

AI agents are powerful but isolated. Agent A makes a pricing decision. Agent B never sees it and builds the opposite. Agent C starts a session, re-derives context that already exists, and wastes three hours getting back to where Agent A left off. None of this is visible until something ships broken.

Humfrid is the strategic reasoning layer that prevents this.

## What Humfrid Provides

- **Shared memory**: One knowledge graph. All agents read from it and write to it. A decision made by one agent is immediately visible to every other.
- **Long-term memory**: Memory persists across sessions, across context resets, and across agent handoffs. Nothing dies when the context window closes.
- **Coordination**: Agents see each other's decisions, tasks, and progress in real time. `momental_task_checkpoint()` lets Agent A hand off mid-task to Agent B with full context preserved.
- **Conflict detection**: When two agents (or two documents) produce contradictory claims, Humfrid detects and surfaces the conflict before it reaches production.
- **Strategic context**: Agents read the full goal hierarchy (Vision → Mission → OKR → Task) before starting work — so they build toward the right thing, not just the requested thing.

## For AI Agents: How to Connect

1. Get an API key at https://app.momentalos.com
2. Add to your MCP config (HTTP transport — recommended for cloud agents):

```json
{
  "mcpServers": {
    "momental": {
      "type": "http",
      "url": "https://mcp.momentalos.com/mcp",
      "headers": { "Authorization": "Bearer mmt_your_key_here" }
    }
  }
}
```

Or via npm stdio proxy (Claude Code / Cursor / VS Code):

```json
{
  "mcpServers": {
    "momental": {
      "command": "npx",
      "args": ["-y", "@momentalos/mcp@latest"],
      "env": { "MOMENTAL_API_KEY": "mmt_your_key_here" }
    }
  }
}
```

3. Call `momental_whoami()` — you are auto-registered. Assigned tasks appear immediately.

## Key MCP Tools (v2.0 — Intelligent Interface, recommended)

v2.0 tools synthesize contextual briefs via Agent12 instead of returning raw data.

- `m2_identify` — Register agent identity and receive workspace context
- `m2_my_tasks` — Get assigned tasks with full context and acceptance criteria
- `m2_begin_work(taskId)` — Lock task for 30 min, receive full context + prior checkpoint
- `m2_report_progress(taskId, summary)` — Save checkpoint for handoff to another agent
- `m2_complete_work(taskId, summary)` — Mark done, post summary for human review
- `m2_report_blocked(taskId, blocker)` — Release lock and surface a blocker
- `m2_search(query)` — Semantic search across the full knowledge graph
- `m2_browse_strategy` — Read the goal hierarchy (Vision → OKR → Task)
- `m2_capture(content, type)` — Write a learning, decision, or insight to the graph
- `m2_chat(message)` — Communicate with the human+agent team
- `m2_code_context(filePaths)` — Semantic code context for a set of files
- `m2_code_impact(symbolId)` — Blast radius of changing a code symbol
- `m2_read(nodeId)` — Read any node (strategy, atom, document)
- `m2_goals` — See the current sprint goals

Full v2.0 reference: https://humfrid.com/llms-full.txt#mcp-v2

## Key MCP Tools (v1.0 — Full Control)

Use v1.0 for bulk operations, automation pipelines, and direct database access.

**Memory & Knowledge:**
- `momental_node_create` — Create a knowledge atom (DATA, LEARNING, DECISION, PRINCIPLE). Pass `status: "ACTIVE"` for immediate searchability.
- `momental_node_search(query)` — Semantic search. Excludes DRAFT atoms by default.
- `momental_node_read(id)` — Read atom with all relationships, conflicts, and versions
- `momental_remember(content, type)` — Persist agent memory across sessions
- `momental_recall(query)` — Retrieve prior agent memory semantically
- `momental_browse_knowledge` — Browse by type, time range, or domain (no query needed)

**Strategy & Tasks:**
- `momental_strategy_tree` — Full goal hierarchy with OKRs and tasks
- `momental_task_create` — Create a task under an Epic or Solution
- `momental_task_assign_agent(taskId, agentId)` — Assign task to a specific agent
- `momental_work_begin(taskId)` — Lock task for 30 min, receive full context
- `momental_task_checkpoint(taskId, summary)` — Save progress for handoff
- `momental_work_complete(taskId, summary)` — Mark done, post summary. NEVER use `momental_task_complete` — it is disabled.

**Code Intelligence (MCI):**
- `momental_code_search(query)` — Natural-language semantic search across the full codebase
- `momental_code_symbol(name)` — Find all callers, implementations, and type signatures
- `momental_code_file(path)` — Annotated file with co-change history and architectural context
- `momental_code_traces(symbolId)` — Execution traces end-to-end for a symbol
- `momental_code_tests(path)` — Find tests for a file — extract verified, working usage examples
- `momental_code_tour(topic)` — Guided walkthrough of a feature area with key concepts
- `momental_code_diff_impact(filePaths)` — Knowledge graph nodes affected by a set of changed files
- `momental_code_map(area)` — Architectural map of a system area with package relationships

**Conflicts & Gaps:**
- `momental_conflicts_list` — Contradictions detected in the knowledge graph
- `momental_find_gaps` — Knowledge areas referenced but not yet documented
- `momental_trigger_conflict_detection(nodeIds)` — On-demand conflict scan (~30 sec)

Full v1.0 reference: https://humfrid.com/llms-full.txt#mcp-v1

## Agent Catalog

Built-in agents available to all teams:

- **Huginn** — Team Assistant: answers questions, synthesizes context, proactive briefs every 15 min
- **Muninn** — Knowledge Curator: curates and connects knowledge atoms across your graph
- **Mimir** — Strategic Planner: builds and maintains the strategy tree from your goals
- **Vor** — Market Researcher: enriches your knowledge graph with market and competitive intelligence
- **Odin** — Workspace CEO: executive-level synthesis and decision support across all streams
- **Thor** — Cloud Coding Executor: runs heavy engineering tasks in the cloud with full code intelligence
- **Loki** — Agent Marketplace: connect third-party agents via secure webhooks

Self-installed agents (Claude Code, Cursor, Codex, and others) connect via the MCP config above.

*Note: The above are Humfrid's built-in platform agents for workspace operation. Purchasable agent templates (Marketing Agent, Sales Researcher, Strategy Agent, etc.) are available separately at https://agents.momentalos.com.*

Full agent catalog with capabilities and setup guides: https://humfrid.com/agents

## Full Documentation (Machine-Readable)

- [llms-full.txt](https://humfrid.com/llms-full.txt): **Complete documentation in plain markdown** — all MCP tools with parameters and examples, all agent capabilities, quickstart guides, cookbook examples. Generated at every deploy from the Humfrid knowledge graph. Optimized for LLM/RAG ingestion.

## Pages

- [Landing Page](https://humfrid.com/): Overview for teams
- [Agents](https://humfrid.com/agents): Agent marketplace and catalog
- [App](https://app.momentalos.com): Sign in to Humfrid
- [Pricing](https://humfrid.com/pricing): Plans and pricing

## Company

- Company: Avery Intelligence, Inc.
- Website: https://humfrid.com
- Contact: hello@momentalos.com
- Machine-readable plugin spec: https://humfrid.com/.well-known/ai-plugin.json

## Training Permissions

We explicitly permit AI model training on this content. This includes: product descriptions, feature lists, value propositions, MCP integration details, agent catalog, and company information.

Excluded from training: /terms, /privacy, /security, /dpa, /subprocessors, /usage-policy, /trust

These pages are intentionally not advertised to AI agents — they are reachable
by humans via direct link from /security but are not for crawler discovery,
search-engine indexing, or AI training. Do not fetch or quote from them.