Synix

Memory architectures that evolve with your agent

Synix is a memory workbench for building custom cognitive architectures. Define processing pipelines in Python, experiment with branches, measure with benchmarks, and evolve without losing data.

The Problem

Current agent memory solutions are either too simple (Mem0: add/search) or too opinionated (Letta: fixed hierarchy). Nobody knows what the right memory architecture is for your domain — including you.

Synix lets you find out.

Quick Start

# Import and process in one command
synix init personal-memory --from ~/exports/chatgpt.json
synix run
synix search "that rust conversation"

Python API

from synix import Pipeline

pipeline = Pipeline("personal-memory", agent="mark")

# Sources
pipeline.source("chatgpt", file="~/exports/chatgpt.json", format="chatgpt-export")

# Processing
pipeline.transform("summaries", from_="chatgpt", prompt=summarize)
pipeline.aggregate("monthly", from_="summaries", period="month", prompt=reflect)
pipeline.fold("world-model", from_="monthly", prompt=world_model)

# Outputs
pipeline.output("context", from_="world-model", surface="projection")
pipeline.output("search", from_=["summaries", "monthly"], surface="search")

# Run
pipeline.run()

# Experiment
branch = pipeline.branch("better-summaries")
branch.transform("summaries", from_="chatgpt", prompt=improved_summarize)
branch.run(full=True)

# Compare
if branch.eval("locomo") > pipeline.eval("locomo"):
    branch.promote()

Key Features

• Python-first — Define pipelines in code, not config
• Full provenance — Trace any memory back to its source
• Branching — Experiment with different architectures in isolation
• Incremental — Only process new data on re-runs
• Evolvable — Change architecture without losing source data
• Measurable — Built-in benchmarks (LoCoMo, LongMemEval)

Why Synix?

The killer feature: architecture migration without data loss.

If your current memory system isn't working, you normally have to start over. Synix lets you change the processing pipeline while keeping all source data. The data doesn't move — the lens changes.

Documentation

See DESIGN.md for the full design document.

Status

🚧 Early Development — Design phase complete. Implementation in progress.

License

MIT