claude-mem

Telemetry: real data edition

Every analytics number claude-mem reports about itself is now real, provider-reported data — plus a new daily install-state snapshot so we can see the actual state of the installed base.

Fixed: the four session_compressed data-quality bugs

• Claude token counts were placeholders. The Agent SDK attaches an early-streaming usage snapshot to assistant messages (output_tokens of ~2–10, regardless of actual output). The session_compressed event is now fired from the SDK result message, which carries the finalized per-turn usage — verified empirically (placeholder said 8, result said 45). Compression ratios for Claude models drop from a nonsensical 6,000–38,000 to the true ~10–100 range.
• cost_usd is now real and populated. Claude: computed from the SDK's cumulative total_cost_usd delta between consecutive turns. OpenRouter: usage.cost + cost_details.upstream_inference_cost (covers BYOK), with usage accounting requested from openrouter.ai only. Gemini reports no cost, so the field stays honestly absent — never estimated.
• Impossible compression ratios (< 1, or exactly 0.0) eliminated. Custom OpenAI-compatible gateways that report suffix-only or one-sided token usage can no longer produce half-real events: usage is now both-sides-or-nothing, ratios require input > 0, and a new endpoint_class property (openrouter | custom) lets dashboards segment gateway-reported data.
• model is never silently missing or wrong. The model that actually served the request (response.model) is stamped instead of the raw configured string, array-typed model settings are normalized, error-path events now carry the model, and unknown is the floor everywhere — non-string values previously vanished in the telemetry scrubber.

New: install-state snapshot

worker_started (start + daily heartbeat) now reports an aggregate snapshot of the local memory DB as person properties: observation/session/summary/project counts, DB file size, install age in days, observations in the last 7/30 days, and days since the last observation. Counts and day-deltas only — never project names, text, or any content. Makes retention, scale, and activity cohorts directly sliceable in analytics.

Also fixed

• The ide person property on worker_started never populated — the lookup queried a legacy table and silently threw on every start since it shipped.
• Epoch math now normalizes legacy seconds-unit rows (a few hundred per install) that would have reported install ages of ~20,000 days.

All new properties are whitelisted in the scrubber, documented at https://docs.claude-mem.ai/telemetry, and shown in the npx claude-mem telemetry consent screen. Telemetry remains anonymous and opt-out (npx claude-mem telemetry disable).

🤖 Generated with Claude Code