VoiceMem vs Arcade AI

Arcade solves a governance problem. Its identity-first architecture authenticates against Okta, Google, or Microsoft Entra ID, delegates per-user permissions so an agent acts as its user and never past its own scope, and enforces policy in the path of the action rather than inside the model. Credentials never leave the runtime, each action produces one audit record naming agent, user, and system streamed to your SIEM, and built-in evals catch hallucinated or destructive calls before execution. The catalog covers 8,000+ permission-aware tools, and recent news extends this: you can now register any third-party remote MCP server — vendor, partner, or internal — and govern its tools from the same control plane with per-user authentication. Deploy cloud, in your VPC, or air-gapped.

VoiceMem solves a memory problem, and specifically in voice. It splits storage into a factual left brain (schemas and entities) and an emotion/persona right brain, runs segmentation, ASR, extraction and graph writes while the user is still speaking, and speculatively prefetches within a 0–300ms window so retrieval begins before the utterance ends. It reports 134ms response time versus Mem0's 1,440ms, 91.2% on LoCoMo with Top-5 memories, and ~430 memory tokens per query turn. Components — memory engine, TTS backend — are swappable, and SessionBuffer purges temporary conversations at session end. There is no overlap: one is about who may act, the other about what is remembered and felt.

Arcade is freemium and metered — the pricing_type tells you the shape but not the numbers, and none are given here, so the honest guidance is that you pay per runtime usage and should size it against your action volume, not your seat count. The value case is build-versus-buy: replacing hand-rolled auth, per-integration credential handling, and audit plumbing with one control point. The costs worth watching are operational rather than listed — VPC or air-gapped deployment carries its own infrastructure and upgrade burden, and the 8,000+ tool catalog only saves you money if you would otherwise have built those integrations. SOC 2, SSO, and RBAC being out of the box matters most if you are being audited; if you aren't, you're paying for assurance you don't yet consume.

VoiceMem is free under Apache-2.0, and that is the entire pricing story — no tiers, no meters, no hosted API. Your real cost is engineering time and local resources: model downloads, warmup, self-hosting, and source-level debugging, plus whatever you spend keeping a v0.0.2 research project running without vendor support or SLAs. The token economics are the explicit design goal: ~430 memory tokens per query turn and Top-K routing to keep context short. Budget accordingly — Arcade is a line item, VoiceMem is a headcount fraction.