Skip to content

Build System

The Synix build system is an open-source pipeline for declarative agent memory.

Status: Shipped. Python, MIT license, 10 releases.

Define memory architecture as a DAG of sources, transforms, and projections. Change a rule, rebuild incrementally. Every artifact is content-addressed with full provenance — you can trace any output back to its sources through SHA256 lineage chains.

Core capabilities:

  • Declarative pipeline definitions with DAG execution
  • Content-addressed artifact storage with provenance tracking
  • Incremental rebuilds (changed layers rebuild, unchanged cache)
  • Hybrid search: keyword (FTS5), semantic, and layered
  • Generic LLM transforms: MapSynthesis, GroupSynthesis, ReduceSynthesis, FoldSynthesis
  • Validation: mutual exclusion, required fields, PII detection, semantic conflict checking
  • MCP server for agent consumption
  • Execution memory protocol (stack+heap model) — 65% token reduction on GPT-5.2

Treating memory as an engineering problem — with versioning, provenance, and lifecycle management — produces better outcomes than flat storage. The build system is essentially MapReduce for memory: repeatable, auditable, diffable.

The build system operates on cold data — conversations and documents that have already been generated. A batch pipeline running every 60 seconds with a 300-second cooldown cannot manage real-time state, coordinate multiple agents, or react to events as they happen.

This limitation motivated the memory architecture.

Full documentation is maintained in the Synix repository: