Skip to content

agent_memory_context.md

Latest commit

 

History

History
25 lines (19 loc) · 1.74 KB

agent_memory_context.md

File metadata and controls

25 lines (19 loc) · 1.74 KB
Raw

PRD: Agent Memory Context (Graph RAG Macros) 🧙🏾‍♂️

Role: Lead Product Manager Persona: Rich Hickey (Focus on declarative data and functional composition over novel stateful engines)

User Story

"As an AI Agent developer, I want a unified, semantically-aware retrieval interface so that my agents can perform Graph RAG and multi-hop memory recall without needing to manually construct complex, multi-join Datalog queries."

Acceptance Criteria

  • Given a semantic intent request (e.g., "Find documents related to concept X connected by relationship Y"),
  • When the request is passed to the rag.gleam query planner/macro layer,
  • Then the request is translated into a pure, optimized AaronDB Datalog AST (combining Bind, ShortestPath, and vector similarity).
  • And the results are returned identically to how standard queries operate, maintaining the EAVT foundation without introducing new storage paradigms.

Technical Constraints

  • Must NOT introduce new storage formats or indexing mechanisms. The macro must operate strictly on top of the existing EAVT and HNSW indices.
  • Must be a pure-functional translation layer: f(SemanticIntent) -> DatalogAST.
  • Should leverage existing native traversals (like ShortestPath and PageRank) inside the AST generation.

UI/UX Notes

  • Developers using AaronDB via Erlang or Gleam will call a new public function rag.semantic_search(intent) rather than building raw q.Query(...) for these specific highly-reusable Graph RAG patterns.
  • AI Agents via MCP will have a simplified semantic_search JSON-RPC endpoint.

Note

PRD Generated. Run /implement (or [/proceed]) to hand off to the Developer Agent to begin implementation according to the Rich Hickey quality standards.