EDM SDK: An Open-Source Tool for Building an Emotional Data Layer for AI Memory Systems

deepadata-edm-sdk is an open-source SDK that extracts emotional data artifacts compliant with the EDM v0.7.0 standard from user content, providing a structured emotional semantic layer for AI memory systems.

Current large language models (LLMs) and AI assistants exhibit impressive language understanding and generation capabilities in conversations, but their handling of "memory" often remains superficial. Most systems simply store dialogue content as raw text or vector embeddings, and rely on semantic similarity matching when retrieval is needed. This "keyword matching"-style retrieval has a fundamental flaw: it can only find content similar in "what was said", but struggles to capture the emotional and meaningful layers of "why it matters".

Imagine you tell your AI assistant a memory about your grandmother. Months later, when you mention "summer" or "old house", you hope the AI can recall that conversation about your grandmother—even if those words didn't appear at the time. Traditional vector retrieval is hard to achieve this because it lacks explicit modeling of emotional importance, memory triggers, and identity associations.

This is the problem the EDM (Emotional Data Metadata) specification aims to solve. EDM does not replace traditional vector retrieval; instead, it adds a "meaning layer"—a structured data layer that explicitly encodes emotional weight, recall triggers, and identity clues.

deepadata-edm-sdk is an open-source implementation of the EDM v0.7.0 specification, providing a complete toolchain for extracting, validating, and encapsulating emotional data artifacts from user content. The SDK's design philosophy can be summarized by three key phrases: Extract rather than infer, structured rather than raw, portable rather than locked-in.

EDM SDK strictly distinguishes between "extraction" and "inference". The SDK uses large language models (supporting Claude, OpenAI, Kimi, etc.) to extract explicitly stated or clearly implied emotional information from text and images, but does not make psychological inferences beyond the input content. For example, it will extract explicit statements like "the user mentioned feeling sad", but will not infer diagnostic conclusions like "the user may have depression".

This design is not only a technical precaution but also a compliance consideration. According to the EU AI Act, emotional data extraction falls into the lower-risk category, while emotional inference systems may face stricter regulatory requirements. EDM SDK ensures its applications comply with relevant regulations through explicit "interpretation constraints".

EDM v0.7.0 defines 10 core domains, covering dimensions from metadata, core emotions, emotional classification to context, salience, motivational states, etc. Each domain has clear field definitions and data types, forming a complete emotional data ontology.

These 10 domains are:

• meta domain: Identity identifiers, data sources, consent basis, and other meta-information
• core domain: Core emotional elements, including anchor, spark, wound, fuel, bridge, echo, and narrative
• constellation domain: Emotional classification and prototype mapping
• milky_way domain: Context anchoring, including related people, places, events
• gravity domain: Salience, weight, and retrieval keywords
• impulse domain: Motivational states, driving forces, and coping mechanisms
• governance domain: Jurisdiction, retention policies, and data subject rights
• telemetry domain: Extraction confidence, model ID, and other technical telemetry
• system domain: Embedding vectors, indexes, and other downstream system fields
• crosswalks domain: Mapping to external taxonomies such as Plutchik's Wheel of Emotions and HMD

This structured design allows emotional data to be accurately retrieved and associated. For example, the system can query "all memories related to 'family' with an emotional weight greater than 0.8" or "find all content that triggers 'nostalgia' emotions".

EDM artifacts use JSON format, which can be freely transferred between different platforms and systems. The SDK also provides integration with deepadata-ddna-tools, supporting the encapsulation of artifacts into signed .ddna envelope format to ensure data integrity and source verifiability.

More importantly, EDM SDK supports "stateless mode", in which all personally identifiable information (such as user ID, related people, location context) is blanked out, leaving only the emotional structure itself. This allows EDM artifacts to be used in privacy-sensitive scenarios, such as temporary conversations or shared analysis.

EDM v0.7.0 introduces the concept of "configuration-aware" extraction, providing three levels of field sets:

Applicable to memory platforms, agent frameworks, and AI assistants. This streamlined configuration includes the most core emotional fields, sufficient to support basic emotional retrieval and association functions while minimizing data collection and storage costs.