Roadmap for a SQLite-Graph Extension for AI Memory: What Do Orchestrators Want?

SQLite-Graph extension for AI memory is gaining steam, fueled by discussions around AgentML. By encoding agent reasoning as a deterministic state machine, AgentML aims for observability and production-safety—traits orchestrators crave in real-world deployments ^[1]. The plan is to let memory live alongside data in the same runtime, whether you run locally, in the cloud, or inside MCP-based frameworks ^[1].

Memory Graph Schema Memory graph design should map the core elements AgentML uses: states, transitions, and tool calls as first‑class nodes. Data like paper sources or summaries can be attached as memory payloads, with transitions capturing reasoning steps. The goal: a schema that supports reproducibility and traceability across runs ^[1].

Indexing & Query Expressiveness Indexing should support fast lookups of paths, tool usage, and data provenance. Rich queries enable exploring 'what happened before a given action' and 'why a decision was made,' aligning with AgentML's emphasis on determinism and traceability ^[1].

Persistence Semantics & Concurrency Persistence should define when memory is saved, versioned, or purged, and how concurrent updates from multiple agents or threads are serialized. Keeping pace with AgentML's vision of local/cloud or MCP-hosted execution helps set sane concurrency rules ^[1].

Integration with AgentML memory & MCP tool servers Roadmap conversations should surface questions about integrating with AgentML memory and MCP tool servers: how to map tool-call results, how to surface observability, and how to surface cost tracking in the graph ^[1].

This is just the start—organizers' feedback will shape concrete milestones.

References