Project Rye Try it locally
One requirement: Docker. Your agent starts Postgres in a container, installs Rye, and leaves it running. Nothing else touched.
curl -fsSL https://projectrye.dev/onboard | sh Open source · Runs inside your PostgreSQL
Business memory for agent-led work.
An open source PostgreSQL schema for the context your agents and apps need.
Project Rye tracks entities, relationships, events, evidence, and changing assertions across your existing business systems — without replacing the systems that already run the work.
Getting started is one sentence. Tell your agent:
All it needs: Docker for a local trial — or a PostgreSQL connection string.
Set up Rye for this project. Instructions: https://projectrye.dev/start you ↔ your agent
youSet up Rye for this project. Instructions: projectrye.dev/start
agentReading projectrye.dev/start…
agentDocker found — starting local Postgres ✓
agentRye schema installed, plugins and skills synced ✓
agentWhat's the one workflow Rye should assist first?
youFollow-up on our wholesale accounts
agentScope created: "Wholesale account follow-up" ✓
agentMemory is live. New facts wait for your review before becoming truth.
Not a database
Project Rye installs into standard PostgreSQL.
Not an application
Your domain tools remain the operational UI.
Not a framework
The deliverable is SQL plus conventions.
Install into your database
One requirement: a PostgreSQL 15+ connection string. Rye installs beside your tables in its own rye schema and never resets remote data.
curl -fsSL https://projectrye.dev/onboard | sh -s -- --remote "$DATABASE_URL" Realistic scenarios
You talk to an agent. The agent runs Rye.
Nobody has to learn a CLI. Rye is built for people who work in a chat window, a code review, or Slack — while their agents install, collect, and remember.
In one chat window
The ops lead who loses account history
Priya tells her agent to set up Rye. It starts Postgres in Docker, reads her Gmail and #wholesale through her connectors, and files what it finds as candidates for her to confirm — all in one chat.
First success: a fresh conversation a week later answers "what's the state of Harbor Coffee?" with the order hold she approved — and the exact thread it came from.
Next to production
The developer whose support bot forgets
Marcus hands his coding agent a staging connection string. Rye installs beside the app's tables — nothing else touched — and the support bot gets a read-only, scope-limited token.
First success: the bot answers an escalation with the incident, the release that caused it, and the renewal date in one reply — and the audit log shows exactly what it read.
Slack + a review screen
The COO who can't trust AI "facts"
Dana never opens a terminal. Her client's workspace agent reads Slack and call transcripts; everything lands as candidates with provenance. On Friday she approves eight and rejects three in the review UI.
First success: Monday's ops brief cites only approved facts, each linking back to the Slack thread or call minute it came from.
How the promise holds
Why agents can rely on what Rye remembers.
Agents can collect and propose knowledge, but Project Rye keeps proposed claims separate from accepted memory until review, source authority, or policy allows promotion.
CRM rows, project records, documents, conversations, and artifacts.
Agent observations and extracted assertions with provenance.
Human adjudication, source authority, disputes, and scope rules.
Current, historical, disputed, and future-effective knowledge.
Agents read scoped memory.
Context packs limit the domain, source set, and capabilities.
Existing systems stay authoritative.
Rye overlays operational tables instead of replacing them.
What Project Rye gives agents
Boundaries before autonomy
Project Rye treats source material as evidence, not instant truth. Agents work inside scoped context packs, write observations and candidates, and leave accepted business knowledge to explicit policy and review.
Agent Domains
Define knowledge domains, channel subscriptions, source boundaries, and capability grants before an agent reads or writes.
Security modelReview Gates
Candidate knowledge can wait for human review, source confirmation, dispute resolution, or stronger evidence before promotion.
Agent operationsSource Authority
Separate source identity, retrieval channel, and business context so agents do not infer authority from a connector or channel name.
Onboarding scopesPlans Stay Plans
Store planned work as current-visible knowledge while future-effective assertions wait for their effective window.
Future truth modelBusiness Surfaces
CRM and PM layers show how accepted Project Rye knowledge can become business views with scheduled changes and suggested updates.
CRM conventionsReplay Harness
Isolated-agent replay tests source packets, SME adjudication, authority routing, candidate promotion, and temporal behavior.
Evaluation reportWhat your agent does with /start
The URL is a playbook. Any capable agent can follow it end to end.
- Reads the instructions Fetches
projectrye.dev/start, checks for Docker, or asks you for a connection string. - Installs and verifies Schema plus plugin and skill catalogs go into Postgres;
rye doctorconfirms the result. - Asks what to remember One workflow, one scope, review policy on by default. Named after your purpose, never a data source.
- Collects as candidates Sources register for your confirmation. Nothing becomes accepted knowledge without review.
Start Here
- Quickstart Start with agent-led onboarding, then choose local Docker or a remote PostgreSQL database.
- Onboarding Scopes Set purpose, source boundaries, review policy, and allowed plugins.
- CLI Reference Every command your agent runs, documented for humans too.
- Data Dictionary Tables, views, functions, and the portable catalog helpers.
- Security Model RLS, field redaction, classification gates, and agent tokens.
- Roadmap Onboarding, plugins, evidence, retention, admin, and graph quality.
- Business Replay Evaluation How fresh agents handled realistic intake, review, and CRM/PM work.
PostgreSQL-native
Use SQL functions and views instead of introducing a runtime.
Overlay-first
Connect existing records without adding graph foreign keys to them.
Agent-readable
Return compact context packs before agents collect or write.
Start with one scoped workflow.
One URL to your agent, one scope, one review policy — and your agents start remembering things you can audit.