Try the reference lab
Understand a failure and its corrected implementation. No installation or connection needed.
Choose a local example or connect your own staging application. Trace checks only the stores and observation window you review.
Understand a failure and its corrected implementation. No installation or connection needed.
Use an existing adapter, bind your app’s functions, or map its staging API routes.
Bring reviewed targets and saved regressions to another machine or CI runner.
Review a bundle ↓This starts an isolated local Mem0 2.2.1 sample. To check your existing Mem0 application, use the Python connector below. This sample is not a connection to your account.
First setup downloads Mem0 and Qdrant client dependencies into a dedicated environment. It needs internet and Python 3.11–3.13; later starts reuse the environment. No paid API key or model inference is used.
Checking local sample status…
Active memories and vectors share one store. Audit history is excluded; no answer cache exists. Trace owns this process and stops it when the engine shuts down.
Bind create, search, inventory and delete functions. Or pass your existing Mem0 Memory instance to the connector.
from synapsoul_trace.sdk import serve_adapter from synapsoul_trace.integrations.mem0 import mem0_adapter # memory is your existing staging Mem0 instance serve_adapter(mem0_adapter(memory), port=8879)
Mem0 connector: tested with Mem0 2.2.1 / Qdrant client 1.18.0. It checks active memories and vector points; it excludes audit history, graph memory and any separate answer cache.
Download Python starter ↓Use the small Node HTTP adapter package with your existing application functions. No runtime dependencies.
import {createTraceServer} from '@synapsoul/trace-adapter';
const server = createTraceServer({
ingest, query, inventory, delete: remove
});
server.listen(8879, '127.0.0.1');Node 20+. Each callback receives the synthetic namespace. Inventory must enumerate the declared stores fully; missing observations are not zero.
Download JavaScript starter ↓Import OpenAPI 3 JSON to choose routes, then review the request and response fields. Trace does not infer what an endpoint deletes or whether an inventory is complete.
Use $namespace in every operation; $id/$text for ingest, $query for query, $ids for delete. Responses use dotted field paths. Inventory needs complete, source_records, vector_records and cache_records. Only an absent, explicitly excluded cache may use {"literal":0}.
trace adapter-serve --mapping trace.mapping.json --port 8879
Start this bridge locally, then connect http://127.0.0.1:8879/trace/v1 in My staging app. Review coverage and test connectivity before running. Supported: static route paths, JSON bodies, scalar GET query parameters, HTTP 200 JSON responses. Path parameters and streaming APIs need an SDK adapter.
Export saved connections and regressions without copying Trace’s evidence database. Bundles contain endpoint URLs, coverage and environment-variable names, never token values. Review their URLs before sharing.
Choose a file to review its destinations. No network calls are made by import.
Existing connections cannot be silently overwritten. Digests detect edits; they are not signatures or proof that a sender is trusted. Import creates no test records. Run and regression replay can create and delete synthetic records.
Completed evidence and connections stay on this computer. An active check finishes before the process exits. Reopen Trace from your launcher to continue.
Run the real local engine against a staging adapter. Only fresh synthetic test records are created and deleted. No application credentials are sent to the public Trace website.
The inventory-only connection check does not create or delete records.
“Checked” means planned API observations; it does not mean independent discovery of every store. The result shows which observations actually completed.
A missing component is never treated as proof of zero records.
Connections are immutable. Use a new name if you change the URL, profile, credentials variable, or coverage.
Verification outcome, coverage, and cleanup are reported separately.
Completed reports survive restart. Resume a running/interrupted job using its ID.
Test what your application forgets.
Keep the evidence. Catch the regression.
Choose a fixture and test sequences.
Results come from observed store contents,
retrieval responses and preserved controls.
API-visible source records, vector records and cached answers after deletion and controlled worker completion. An unrelated record must survive.
Disk or backup erasure, model unlearning, historical chat deletion, production coverage, or independently reproduced upstream vulnerabilities.
Reports are append-only through Trace. Saved test definitions are content-addressed. Local file access is not a security boundary.
Confirmed outcomes inform which test runs first. This is a transparent history heuristic—not model training or evidence of generalization.
Calculated per backend over the latest 500 reports. Inconclusive tests do not count as successes. Fixture mode influences these observations; this is not a production risk estimate. Test assertions are never relaxed.
Connect an MCP-compatible client to the same local engine. No model subscription is required to use Trace's CLI or dashboard.
Source install: install the optional MCP extra first (the Windows app includes it): python -m pip install -e ".[mcp]". This configuration uses the Python executable running this dashboard. Restart your client after adding it.