Doplex Labs builds systems that help AI agents learn from real work with evidence and developer control. Belay is our first product.
The same lesson should not need to be taught twice.
Coding agents can solve increasingly difficult tasks, but practical learning is often trapped inside one session. A developer corrects an approach, explains a project constraint, establishes the right verification command, and confirms the final solution. The next session can begin with none of that judgment.
The result is familiar: developers repeat corrections, agents rediscover failed approaches, and instruction files grow without evidence that every added rule actually helps.
A developer should not need to teach the same important lesson twice.
Belay is a private experience layer for coding agents.
Belay identifies useful lessons from completed work, keeps the supporting evidence on the developer's machine, asks for approval, and prepares the lesson for a future agent when it is relevant.
- An agent attempts a task.
- The developer corrects the approach or establishes a rule.
- The final solution is checked with tests or another concrete result.
- Belay links the lesson to the original work.
- The developer reviews and approves it.
- A future agent receives the lesson while working on a related task.
Belay separates what happened from what an AI thinks the lesson might be and what the developer approved. We call this verified experience: a reusable lesson that remains connected to the work and evidence that produced it.
Useful experience, not another pile of context.
Memory asks what information should be available later. Belay asks which past lesson matters now, where it applies, who approved it, and how the agent should verify its work.
The value is relevance, not volume. Belay carries forward specific, evidence-backed lessons instead of copying entire histories into every future session.
What we measured.
We evaluated Belay with Claude Code and Codex on new software and machine-learning tasks. Each task tested whether a lesson from earlier work could help in a new situation while remaining appropriately scoped.
Tasks completed when an earlier lesson was useful. First evaluation across Claude Code and Codex.
The result demonstrates the potential for carefully selected lessons from prior work to help future sessions. Broader evaluation across agents, tasks, and real development workflows is underway. Doplex Labs will publish the full methodology and agent-level results in a research preprint.
Private by design.
Belay stores coding-session data in an encrypted local database and does not upload it to Doplex Labs. It requires no Belay account and ships no AI model.
Optional interpretation uses the developer's installed Claude Code or Codex under their existing provider configuration. Proposed guidance remains inactive until the developer approves it.
Where Belay is going.
Our next work will make each lesson's scope more precise, adapt guidance to the agent and project, and measure repeated corrections, review effort, and time to useful progress in real development workflows.
Belay will be measured by outcomes, not by the number of memories it stores.