Telemetry¶

Membrick telemetry is local operational data used to understand query behavior and provider activity. It is separate from indexed memory and session archives. Aggregate telemetry is bounded. Detailed MCP tool history retains complete arguments and responses.

Telemetry streams¶

Membrick currently has four related streams:

HTTP query telemetry is process-local. Provider MCP processes persist their query aggregates so the dashboard can observe queries handled outside the HTTP server process. The query aggregates are project-wide and do not attribute a request to a provider. Detailed history is recorded after every provider tools/call, including search, memory reads and writes, memory controls, and update actions. The Claude, Codex, Gemini CLI, AGY, Muse, OpenClaw, Hermes, and Roo runtime adapters all use the same recorder.

Query aggregates¶

Queries are grouped into five-second buckets. The most recent 120 buckets are kept, giving a rolling ten-minute window. Each bucket tracks request count, error count, empty-result count, and a latency histogram. The dashboard derives request rate and latency summaries from these aggregates. Query strings, session IDs, and document contents are not stored in this aggregate stream.

The detailed MCP tool history is an append-only JSON Lines file with no retention, entry-count, or payload-size limit. Each entry includes provider, tool name, complete arguments, timestamp, elapsed time, success or error status, and the complete response returned by the resident handler. Session IDs and content are retained as supplied. On Unix, the file is created with owner-only permissions. Entries can contain private project information and will grow until manually removed, so treat this file as sensitive and account for its disk usage. It is an operational archive, not a semantic memory document, and is not included in normal memory search results.

Provider lifecycle events¶

Provider events are kept in a separate JSON ledger for each provider. Each ledger retains at most the latest 500 events. The event schema can include:

• event name, category, and timestamp
• a one-way session key rather than the provider's raw session ID
• bounded agent, turn, and tool metadata when the provider supplies it
• redacted, bounded previews for selected prompt or tool activity
• retrieval timing and result counts when available
• normalized token counts when available

Lifecycle telemetry is written independently of session recording. Turning off session recording prevents full event archives from being written; it does not by itself disable operational lifecycle telemetry. Full session content is not copied wholesale into the telemetry ledger.

Provider coverage¶

Coverage depends on what each provider exposes and which integration features are compiled:

“Yes” means the code path is instrumented; it does not mean telemetry exists for a provider in a particular workspace. A provider remains not_observed until it emits an event. Token usage may be unavailable when the host does not include usage data in its hook payload or transcript.

The repository also contains a separate optional Python Hermes hooks plugin for memory capture. That plugin is not connected to the provider telemetry ledger described here, so it does not make Hermes lifecycle telemetry appear on the dashboard.

Dashboard and API¶

The local dashboard reads query and provider telemetry through these endpoints:

See HTTP server for starting the dashboard and configuring its project root and authentication.

File watcher changes¶

Workspace file diffs are not operational telemetry. The watcher captures bounded before-and-after snapshots and persists eligible diffs as memory SourceDocuments. It does not send file contents or diffs to the telemetry ledgers. This keeps source content in the memory store and operational metrics in the telemetry path.

Content capture skips .env files, private-key files, and standard credential locations such as .ssh and .aws. Known credential values and sensitive assignments are redacted from captured diff lines. This is best-effort filtering; custom secret formats and arbitrary secret filenames may need additional ignore rules. Existing stored documents are not retroactively sanitized.

On Unix, snapshot reads reject symlinks in every path component and use held directory descriptors to avoid check/open races. Reads are bounded to 256 KiB. Automatic snapshots and rooagi_runtime file payload references are unavailable on non-Unix targets until an equivalent safe reader is implemented. Watcher invalidation and inline hook payloads continue to work there.

Implementation references¶

• src/telemetry.rs implements query aggregates, provider event ledgers, and MCP call history.
• src/integrations/session_recording.rs normalizes provider events and token usage, and separates operational telemetry from full session recording.
• src/pipeline/watcher.rs captures bounded workspace file snapshots for diffs.