Build AI Agents With Memory That Shows Its Work
Open-source governed memory and workflow receipts for AI builders
Capture what agents did, decide what becomes knowledge, and run repeatable workflows with evidence instead of vibes.
Start Here · Quick Start · What Can You Build? · Architecture · Workflow Library · Deployment · Library · Catalog
Why Allura?
AI agents forget. Sessions end, context evaporates, and your team's hard-won knowledge disappears into the void.
Allura gives builders a self-hosted memory and governance kernel for agent systems. It is not a black box that silently decides what matters. Every memory starts as traceable evidence, then moves through scoring, review, and promotion before it becomes long-term knowledge.
Use Allura when you are building agents that need to:
- remember decisions across sessions;
- show where an answer or decision came from;
- separate raw traces from approved knowledge;
- route work through repeatable story, review, architecture, or run workflows;
- produce receipts before claiming work is Done.
Allura is for builders who want:
- 🔍 Inspectability — trace what was recorded, when, and why it was promoted
- 🏛️ Governance — approval gates between raw capture and long-term knowledge
- 🔒 Self-hosting — your data, your infrastructure, your rules
- 🧩 MCP-native — plug into Claude, Cursor, OpenCode, or any MCP-compatible agent
- 🧭 Workflow receipts — story, review, architecture, and run records with evidence gates
Start Here
New to Allura? Pick your path:
Quick verification after install:
docker compose up -d
curl http://localhost:3201/health
What Can You Build?
Allura is a foundation for builders making their own governed agent systems.
| Builder goal | Allura gives you |
|---|
| Persistent agent memory | MCP tools for add/search/get/list/delete with tenant isolation |
| Approved knowledge base | Raw PostgreSQL traces plus promotion-gated Neo4j knowledge |
| Agent cowork handoffs | Shared handoff packets and validation receipts for Claude, Codex, OpenCode, Cursor, and other runtimes |
| Evidence-gated workflows | Story, review, architecture, product-intake, retrospective, and RunRecord templates |
| Audit-friendly automation | Append-only events, actor identity, source metadata, and explicit approval points |
| Your own methodology library | A public library/ pattern you can copy, adapt, and extend |
If Archon is primarily a workflow runner, and Babysitter is primarily process enforcement, Allura is the governed memory and evidence layer that lets your workflows remember and justify themselves.
Workflow Library
The library/ directory is the public chooser for builders. Start there when you want a repeatable process rather than only raw memory calls.
Choose a workflow → story, review, architecture, product intake, RunRecord
Choose an adapter → Claude, Codex, OpenCode, Cursor, or MCP HTTP
Choose governance level → local receipts, review-gated promotion, or full evidence packet
Run it → produce memory writes, validation evidence, and a closeout receipt
