The Assembly
One artifact. Complete agent specification. Enforceable contract.
The Assembly is the central artifact produced by Mimory's context engine. It captures everything an agent needs to operate: its persona, tools, permissions, relationships, and boundaries. More than just configuration, the Assembly serves as an enforceable contract that projects directly into IAM policies, MCP Gateway rules, and agent execution frameworks.
What is an Assembly?
An Assembly is the complete, signed specification that defines an agent's identity, capabilities, and constraints at a moment in time.
Unlike static configuration files, assemblies are dynamically synthesized from your context maps, binding rules, and real-time relationships. Each assembly is unique to a specific role, collection, and moment.
Every assembly is cryptographically signed and logged with full traceability of how it was arrived at, making it both machine-readable and auditable.
Assembly Structure
General Info
Basic metadata about the assembly itself: assembly_id and assembled_at timestamp.
These identifiers enable precise audit trails and allow you to reference specific assembly versions in logs and compliance reports.
Character
The agent's persona and behavior configuration. Includes the system prompt, welcome message, pre/post-tool prompts, and optional tuning parameters like model selection and temperature.
"name": "John",
"pre_tool_prompt": "You are John, a kind assistant...",
"config": { "model": "gpt-5.1-2025-11-13" }
}
Tools
List of Model Context Protocol servers the agent can connect to. Each MCP entry includes the allowed tools, parameter constraints (focus), permission mappings, and OAuth scope requirements.
- •tool_permission_map: Maps tools to required IAM permissions
- •focus: Constrains parameter values (e.g., only patient P001)
- •scope_map: OAuth resource scopes to request
- •guidelet: Usage hints for the agent
Other APIs and CLIs can eventually be exposed as toolchains, similar to MCPs, but this requires certain base tools in the agent execution environment.
Subagents
Local subagent descriptions that run within the same execution context. Similar to the agents array in Claude Code or other coding assistants, these define specialized helpers the main agent can spawn.
Subagents inherit the parent assembly's security boundaries while operating with their own focused capabilities.
Skills
Standardized, reusable skill definitions the agent can invoke. Skills encapsulate common workflow patterns with associated tools.
Skills promote consistency across agents and simplify capability composition.
Agents
Remote agents that can be invoked for specialized tasks using A2A protocols. Each agent entry defines its endpoint, authentication, initialization parameters, and permission requirements.
Example: An "Insurance Checker" agent that verifies procedure coverage before scheduling.
MIM Graph
A compact map of the agent's operational world, describing nodes (entities) and edges (relationships) that can be injected into the agent's main prompt. This gives the agent situational awareness of its environment.
Nodes: Patients, projects, rooms, equipment
Edges: "P001 is the patient in Room 1", "Station 2 operating in Room 4", "P003 is a sibling of P002"
Full Traceability
Every assembly is logged with complete provenance. You can trace back through binding rules, asset definitions, and sensing events to understand exactly why an agent received specific capabilities.
The assembly includes a cryptographic signature that guarantees integrity and enables verification that the agent operated under the correct constraints.
"algorithm": "RSA-PSS-SHA256",
"signature": "L9Uur2TY...",
"timestamp": "1763585970",
"kid": "c2b54b5f-..."
}
One Artifact, Multiple Projections
The Assembly isn't just agent configuration, it's a universal contract that projects into every layer of your stack.
Agent Runtime
Projects into any coding agent framework: Claude Code, OpenAI Assistants, LangChain, or custom runtimes.
IAM & Gateway
Tool restrictions and parameter constraints enforce at the gateway level before requests reach your services.
OAuth Scopes
Scope maps determine exactly which OAuth scopes to request from each authorization server.
Evolving with Open Standards
The Assembly format connects with emerging agent ecosystem standards for subagents, skills, and interoperability.
Subagents are newer and emerging primarily with Claude Code, but variations appearin in other frameworks such as LangChain DeepAgents.
Boundaries by design. Verifiable by default.
Build agents that carry their contracts with them: auditable, enforceable, and always accountable.