Tenant-scoped event memory
Typed events, source links, embeddings, vector search, and Cypher traversal sit behind one API. Memory is evidence, not a loose note store.
Platform
One trace for memory, runs, and replay.
TraceBreak is the operational layer for agentic software. It records every step as linked events, keeps the run loop inspectable, and replays the same chain against counterfactuals when behavior changes.
Product model
Every decision stays attached to its evidence.
Instead of shipping memory, execution, and debugging as separate products, TraceBreak treats them as one event lineage. That gives operators a durable path from final output back to the exact source events and model decisions that produced it.
Capabilities
The three pieces that make a run debuggable.
TraceBreak is intentionally narrow: capture what happened, preserve why it happened, and replay the path that matters.
Agent loop with an audit trail
Each plan, tool call, tool output, and stop reason is written as it happens. The run can be inspected later without reconstructing logs.
Counterfactual debugging
Re-run the same lineage against a different prompt, model, or state snapshot and compare where the result diverges.
Workflow
From incident to fix without losing the path.
Record the run
TraceBreak writes linked events from the first goal through the final stop reason.
Open the failed event
Start from the output, error, or human escalation and walk backward through source events.
Replay the lineage
Hold the evidence path constant while changing the model, prompt, or state under test.
Compare the diff
Find the first changed decision and attach the replay result to the fix.
Deployment shape
Self-host first, cloud-ready later.
The current stack is built around a Rust service boundary, explicit tenant isolation, and simple HTTP APIs. Teams can start with a single host and grow into managed deployment as the product matures.
OpenAI-compatible LLMs
Use providers that expose the familiar chat and tool-call shape. Native provider adapters can come later.
FFS-backed storage
Graph, vector, columnar, and provenance live in the storage engine; TraceBreak adds tenant and operator surfaces.
Operator visibility
Structured logs, request correlation, metrics, and replay artifacts make incidents easier to explain.
Next step
Bring a real failed run.
A demo is strongest with a run you already had to debug manually.