Lavra -- Compound Engineering & Memory for Agents

Compound Engineering Workflows with Memory for AI Agents

Build better and faster

Lavra gives your coding agents persistent memory, multi-agent compound engineering workflows, and knowledge that compounds across sessions.

Get Started View on GitHub
Claude Code OpenCode Gemini CLI Cortex Code

AI agents have no consistent memory.

  1. Agents forget what they learned last session.
  2. They are overeager to jump to implementation with basic plans that leave a lot of leeway for agents to come up with wrong or suboptimal implementations. [3]
  3. Basic or no research into patterns or current best practices, going off their usually outdated training data.
  4. Code reviews are done only when asked, with no consistency.
  5. They tend to repeat (wrong) code many times over, for lack of guardrails. [1]
  6. Working with only the main agent (the one you chat with) is less optimal than with subagents given well-defined tasks. [2]
  7. Shipping is five manual steps.
An AI agent with no memory, starting from scratch every session

References

  1. [1] Mario Zechner on AI agent memory loss
  2. [2] Simon Willison -- Agentic Engineering Patterns: Subagents
  3. [2] Gabriella Gonzales -- A sufficiently detailed spec is code

Four commands in the main pipeline.

From idea to shipped code. Power users can tap in directly to related commands and skills.

/lavra-design

What to build. Vision & success criteria. Create beads issues with domain-matched research agents that match the plan
/lavra-work

Implement with parallel subagents, automatic knowledge recall feeding correct context, multi-agent reviews and fixes
/lavra-qa optional

Browser-based testing from the user's perspective
/lavra-ship

Automated shipping with tests, security scans, and PR creation
30
Agents
27
Commands
15
Skills
4
Platforms

Compounding memory

Your agents learn from every session. Knowledge captured today is automatically recalled when relevant.

March 2025

$ /lavra-work BD-042

Implementing OAuth integration...

Review passed.

 

[auto-captured]

LEARNED: Redirect URI must match exactly

PATTERN: Validate callbacks server-side

DECISION: Use PKCE for public clients

Three weeks later

$ /lavra-work BD-089

 

[auto-recall] 3 entries matched

LEARNED: Redirect URI must match exactly

PATTERN: Validate callbacks server-side

DECISION: Use PKCE for public clients

 

Starting with context...

Hit the same problem next month? Your agent already knows the fix.

See it in action

Get started

npx @lavralabs/lavra@latest
Claude Code OpenCode Gemini CLI Cortex Code