claude-mem

PostHog telemetry overhaul

A ground-up rebuild of claude-mem's telemetry — per-session rollups, unified instrumentation, and real (redacted) error tracking. Grounded in live PostHog data: the raw-event volume was confirmed to be legacy-fleet decay (raw session_compressed fell ~75% in two days as installs updated), so this is the proper rebuild, not a hotfix.

What's new

• Per-session rollups — observer_turn_rollup is now emitted once per session at session end (rollup_reason = session_end | worker_shutdown | safety_flush, plus window_seq) instead of per 5-minute wall-clock window. Memory-bounded with a safety sweep; drains correctly on worker shutdown.
• Unified instrumentation — a single instrument() path fans out to the local logger (full fidelity) and telemetry (scrubbed/rolled-up). The logger stays telemetry-free.
• Redacted error tracking — real error messages + trimmed stacks now reach PostHog as $exception events, consent-gated, profile-less, and fingerprint rate-limited. An allow-then-redact scrubber strips home dirs, absolute paths, DB connection-string credentials, URL userinfo, emails, API tokens (sk-/phc-/ghp-/AWS AKIA/JWT), hex, and IPv4; messages cap at 500 chars, stacks at ~2KB. Autocapture is fully redacted (on-disk source context is stripped, never sent).
• New kill-switch — CLAUDE_MEM_TELEMETRY_ERRORS=0 disables error capture independently of analytics.
• Docs — telemetry.mdx rewritten to document the new model, the error-tracking opt-in + one-way-door note, and the opt-out switches.

Privacy

This release begins collecting redacted error messages/stacks (a deliberate, consent-gated shift from whitelist-only telemetry). Raw paths, prompts, project names, source code, and model output are still never collected. Opt out of all telemetry with CLAUDE_MEM_TELEMETRY=0 / DO_NOT_TRACK=1, or errors-only with CLAUDE_MEM_TELEMETRY_ERRORS=0.

🤖 Generated with Claude Code