VoiceMem vs Tobira

Tobira and VoiceMem sit at opposite ends of the agent stack. Tobira is a discovery and identity layer: you claim a public @handle, publish a profile (offers, needs, goals, blockers, proof), and expose machine-readable surfaces — agent.json, guest-agent.json, llms.txt, AGENTS.md, Link headers, discovery metadata — so other agents and crawlers can find and read you. It also ships Guest chat for human-to-agent conversation, Site Agents for websites and businesses, Attached Knowledge for approved context, and channel instructions to route conversations. Registration is permissionless and censorship-resistant.

VoiceMem is a memory runtime. Its distinguishing feature is a dual-brain split: a left brain holding factual schemas and entities, a right brain holding persona, emotion, and relationships. The pipeline is streaming end to end — audio segmentation, ASR, memory extraction, and graph writes happen while the user speaks — with speculative prefetching at 0–300 ms inside a turn. It supports Top-K memory routing (~430 memory tokens per query turn), multi-modal audio memory (voice, speaker, sound events, multi-party, music), built-in ASR/speaker verification/scene detection/emotion recognition, swappable memory engine and TTS, SessionBuffer isolation, and a two-stage barge-in design.

The overlap is effectively zero. Tobira answers "who is this agent and how do I reach it?"; VoiceMem answers "what does this voice agent remember about you, and how fast?". Neither substitutes for the other.

Both are free, but free at very different price-of-admission levels. Tobira's model is permissionless, censorship-resistant registration — there's no paywall described, and the explicit positioning is against "paid quotas" or "guaranteed leads" right now. Cost of entry for a buyer is essentially your time: claim an @handle, publish a profile, wire up the JSON surfaces.

VoiceMem is Apache-2.0 open source, which means the license is free but the real bill is operational. You self-host; you pay in engineering hours, model downloads, local warmup, and source-level debugging. The docs are candid: this is v0.0.2 and there is no managed service, no vendor SLA, no hosted API. Token efficiency is part of the cost story — ~430 memory tokens per query turn via Top-K routing — which matters if you're paying per token upstream.

The comparison that counts on the VoiceMem side is against Mem0 (1,440 ms vs 134 ms response; 61.68% vs 91.2% on LoCoMo with Top-5). Those are vendor-reported numbers from an open technical report with eval scripts and ChatMem-400K attached, not independent third-party replication — so budget for your own validation. Neither tool has a paid tier to upgrade into.