Performance

Cost is an architecture property, not a tuning accident. The whole pipeline is shaped so ordinary operation stays fast and cheap, with model spend concentrated where judgment is genuinely needed.

Where the money does not go

• The gate admits a fraction of traffic before any model sees it. "Hey" costs nothing; newsletters cost a table lookup.
• Identity is dictionary lookups at ingest — Contacts, sender policy, nicknames — never per-line model calls.
• Daytime collection and brief rendering make zero model calls. Freshness hints are deterministic; only requested evidence interpretation spends.
• Prompt caching does real work: the shared prefix (event window, to-dos, standing) is identical across every bundle call in a run.

Where it does go

• Propose: one call per packed group per pass (up to MEMCAL_PACK_BUNDLES, default 6, bundles per call; a typical day is 5–15 bundles total), bounded by MEMCAL_ITEM_BUDGET and MEMCAL_PACK_TOKENS. Parallelism caps at MEMCAL_MAX_PARALLEL (default 8).
• Sweep: one cheap call re-reading resulting state, not the day's traffic.
• Merge arbitration: only genuinely ambiguous near-horizon conflicts.
• memcal who --resolve: one call over the whole identity picture, asked for explicitly.

memcal stats reports post-gate volume and run history — instrument before optimizing. The gate's effectiveness is the entire cost story, and no estimate is worth more than a week of real numbers.

Latency shape

Prepared answers read the brief: immediate. Flagged items add one bounded activity read. Nothing in the turn path waits on extraction, consolidation, or history-wide search. Response latency, source reads, driver and provider model calls, and nightly processing cost are recorded separately, so a slow correct answer and an immediate correct answer are never confused. See Grading.

Backends

Backend	Default model	Authentication
Codex programmatic mode (default)	gpt-5.6-luna	Existing Codex login
Claude Code programmatic mode	claude-sonnet-5	Existing Claude Code login
Antigravity programmatic mode	gemini-3.8-flash-high	Existing Antigravity login
OpenRouter	openai/gpt-5.6-luna	OpenRouter API key

The three CLI backends run as one-shot structured completions. Claude Code uses print mode without persistent sessions or tools. Codex uses ephemeral exec sessions with a read-only sandbox and approvals disabled. Antigravity (agy) uses print mode with the sandbox on and slash commands disabled, and asks for structured output with --json-schema. All three replay staged extraction turns explicitly, so a pass does not depend on hidden session history. Prefer a flash model on Antigravity: its print mode wraps every prompt in an agent preamble that bills tens of thousands more input tokens per call — quota rather than money on a subscription, but the reason to prefer flash here. Antigravity model names carry their own reasoning budget (gemini-3.8-flash-high and gemini-3.8-flash-low are separate selections), so memcal leaves --effort alone for those and sets it only for a model that does not state one; agy models lists what the login can reach. Antigravity returns SUCCESS with an empty response often enough to notice — roughly one call in five in a small sample. Memcal treats that as a failed call rather than an empty answer, so it shows up in a pass's failed count and the bundle stays available. See Configuration.