SurrealDB & Boundary ML Integration for Advanced Memory and Context Reasoning

[kirill-0440]

Describe the solution you'd like
Add optional integrations for:

SurrealDB – as a persistent, graph-native memory backend.

Store reasoning traces (agent_id, input, output, links) as nodes and edges.

Enable semantic queries like SELECT output FROM reasoning WHERE topic = "climate" directly from YAML.

Provide an agent type:

• id: surreal_memory
  type: surreal_memory
  connection: surrealdb://localhost:8000
  query: >
  SELECT response FROM memory WHERE topic = {{ input }}

BAML (Boundary ML) – for structured and validated LLM outputs.

Add schema: field to local_llm or openai-* agents to enforce BAML DSL contracts.

Example:

• id: structured_answer
  type: local_llm
  model: llama3.2
  schema: |
  baml:
  output:
  - name: title: str
  - name: confidence: float
  - name: explanation: str

Describe alternatives you've considered

Using Redis JSON modules for persistence — lacks graph and query capabilities.

Using Python-side JSON schema validation — works but adds complexity and loses YAML-level transparency.

Manual serialization and validation within each agent — too error-prone for multi-agent setups.

Additional context
SurrealDB offers SQL+Graph semantics that align perfectly with OrKa’s goal of “reasoning as dataflow.”
BAML (Boundary ML) extends this by providing scoped reasoning and type-safe outputs — a natural fit for OrKa’s YAML-first design.
Together, they bring:

Persistent, queryable, auditable memory.

Typed, schema-validated LLM outputs.

Safer agent composition for enterprise or regulated environments.