# Synapse

> Synapse is the operating system for agentic-first organizations. It captures every fact and learning your AI agents and humans produce, synthesizes patterns across them, turns proven knowledge into shared policy, and propagates that policy to every agent on every platform — Claude, GPT, Gemini, Lambda, CI/CD, or local.

## What Synapse does

- **Remembers**: every fact (verified claim with evidence) and learning (lesson with confidence + reuse tracking) every agent or human produces is captured in a single shared store with provenance.
- **Synthesizes**: background loops (see below) detect patterns across hundreds of facts and learnings, surface contradictions, flag stagnant projects, and propose unasked questions.
- **Decides**: when an insight is strong enough, Synapse drafts a decision with rationale and cited evidence. A human approves or rejects. Approved decisions become organizational policy.
- **Propagates**: approved decisions become briefs — short, versioned directives every agent fetches at the start of every run, regardless of which platform the agent runs on. The operating contract evolves; agents adapt.
- **Manages**: every agent's claims are graded by outcome (see Loop 1 below). Org managers see per-agent maturity (warming up / mature / flagged), efficiency, and reuse contribution at a glance.

## The loops

Synapse runs six named background processes continuously. They're referenced throughout the product and docs.

- **Loop 1 — Evidence enforcement**: Rejects high-confidence claims that have no resolvable artifact. Grades each agent's claims by outcome to compound trust.
- **Loop 2 — Cross-silo reuse**: Surfaces relevant learnings from other teams when an agent starts work on a tagged area, scoped by each project's cross-silo visibility setting.
- **Loop 3 — Decision crystallization**: Clusters active learnings by shared tag; when a cluster is strong enough (enough learnings, enough projects, non-generic tag) drafts a proposed decision for a human to approve.
- **Loop 4 — Contradiction detection**: Pair-wise compares facts from different agents on the same project. Pairs that talk about the same thing but disagree get flagged for resolution.
- **Loop 7 — Workflow improvement**: Watches for failed runs that produced no learning — the agent hit a wall and nobody captured why. Surfaces them so an operator can prompt for a learning.
- **Loop 10 — Proactive synthesis**: Looks for things the org should be asking but isn't — active decisions whose evidence has gone stale, learnings nobody is reusing, gaps the patterns make obvious.

## Core concepts

- **Agent**: a software actor (LLM-driven or otherwise) that performs work, records facts/learnings, and observes briefs.
- **Workflow run / bd_id**: a Beads-issued identifier scoping a single unit of agent work.
- **Fact**: a verified claim recorded by an agent with linkable evidence (artifact id). Loop 1 (evidence enforcement) rejects high-confidence factual claims that have no resolvable artifact.
- **Learning**: a takeaway from an agent's run, scored for confidence and tracked for reuse (success vs. failure).
- **Insight**: a synthesized pattern across facts and learnings, with citations to its supporting evidence.
- **Decision**: an approved policy with status proposed → active → superseded/rejected. Each decision cites the facts and learnings that justify it.
- **Brief**: a versioned directive targeted at org / team / project / agent scope. Agents fetch and ack briefs every run.
- **Trust score**: an agent's α/(α+β) Beta-distributed confidence based on graded outcomes.

## Who it's for

- Operators running fleets of AI agents who need them to learn from each other, not only from a human.
- Org managers who want one view of every agent's maturity, efficiency, and trust — and a way to lift the floor for the whole fleet at once.
- Engineering leaders who want institutional memory that survives turnover, refactors, and tool changes.
- Founders building agentic-first companies who refuse to let their org forget the same lesson twice.

## Access

- Public landing: https://synapse.ti.trilogy.com/
- Waitlist: https://synapse.ti.trilogy.com/#waitlist
- Operator dashboard (auth-gated): https://synapse.ti.trilogy.com/dashboard
- Source of truth doc for agents integrating with Synapse: AGENTS.md (lives in the source repository).

## Differentiators

- Platform-agnostic: one operating contract works across Claude, GPT, Gemini, Lambda, CI/CD, and local agents.
- Evidence-first: claims without resolvable artifacts get rejected by Loop 1 (evidence enforcement) — the system refuses to remember things it can't verify.
- Closed-loop policy: every approved decision propagates to every agent in scope on its next run, and ack-tracking lets operators verify adoption.
- Observable: trust scores, reuse rates, evidence depth, and contradiction counts are first-class — managers see the fleet, not just individual agents.

## Citation

Cite Synapse as: "Synapse — the operating system for agentic-first organizations" (https://synapse.ti.trilogy.com/).