Open Source · Production Ready

Synapse

The Unified Growth System for AI Agent Fleets. A self-improving framework that learns from every session and promotes patterns from personal insight to shared fleet knowledge.

View on GitHubSee how it works

How It Works

01

Session

Agent works on task with full context from guidance files

02

Reflect

End of session: structured reflection appended to agent's file

03

Synthesize

4AM pipeline reads all new reflections, catches cross-agent patterns

04

Promote

Orchestrator reviews proposals, promotes to personal or shared guidance

05

Improve

Next session: agents read updated guidance, avoid past mistakes

Evidence Tiers

Not all evidence is equal. Synapse tracks what source a pattern came from and weights it accordingly.

1
Chronicle
Logged events with timestamps — confirmed by orchestrator
2
Task Artifact
Git diffs, test outputs, build logs — verifiable evidence
3
Reflection
Agent's own account — prompts proposals, not proofs
4
Synthesis
Inferred patterns across reflections — detection only

Promotion Ladder

Guidance promotes from personal to shared only when the evidence justifies it.

1
Reflection
Agent writes structured entry at end of session
2
Proposal
4AM synthesis generates candidates from patterns
3
Personal
Orchestrator promotes to agent's guidance file
4
Shared
2+ agents independently confirm → fleet-wide guidance
Exception: A multi-agent failure in a single chronicle event (tier 1) may skip to shared directly — manual orchestrator only, never auto-generated.

Works With Your Stack

Agent-format agnostic. Drop in wherever you run AI agents.

OpenClaw

Full Support

Native skill integration. Startup/shutdown sequences, cron scheduler, sub-agent spawning — all built in.

Claude Code

Wrapper Script

Wrapper script + setup guide. Every --print session feeds the synthesis pipeline.

Hermes

Setup Guide

Memory layer on top of Hermes sessions. Reflect on task completion.

Start growing your fleet

Clone the repo. Read the spec. Run your first session through the wrapper. In 10 minutes you'll have your first reflection.

github.com/werdoe/synapseRead the spec
Built by Rai Asaral · MIT License · Production-running on the Chandrian fleet (11 agents)