Companions, coaches, tutors, eldercare, hardware that lives in users’ homes. Voice agents that need to remember, evolve, and stay in character for months. Bring your STT, LLM, and TTS — Hyponema holds the relationship.
How it actually works
Most voice agents start every conversation from zero. Yours pick up where the last one ended, with the right tone, the open thread surfaced, and the topics that should rest.
I'm not great today. The week's been heavy.
Take your time. I'm here.
Heavy session. The week has been rough on her.
Don't bring up work for the next few days.
She left the call heavy. The agent flags work as a tender topic for the next few days.
Be gentle. Last call was hard.
Ask about the weekend, not work.
Hi. How has the rest of the week landed?
Work stays out of the conversation. It is resting until Friday.
The agent opens differently. Because it knows.
her baseline lifted across 11 conversations
The agent walked with them.
Across eleven conversations, the agent moved with her. The relationship has weight.
Relational memory
Hyponema holds the relationship itself: the tone of last week’s call, the promise from a month ago, the topic that should rest until Friday.
How the engine works0.4·recency + 0.3·frequency + 0.3·relevance. Health, family, identity carry floors so they never decay below 0.5 even after months of silence.
Contradictions create explicit linked records with reasons. Old facts stay queryable. Consolidation merges near-duplicates at cosine ≥ 0.75.
Longitudinal storylines detected post-call (cosine ≥ 0.65). Lifecycle: ACTIVE → DORMANT (after 60 d silence) → RECURRING.
14-day rolling vs 60-day baseline drift. Pure tabular computation, no LLM call. Topic cooldowns auto-suppress sensitive users for 3 days after a NEGATIVE / CONCERNED episode.
"Where did we leave off" snapshot — last episode tone, top open question, outstanding promises, recommended resumption tone. Injected at priority 0 of the system prompt.
Drift detector pre-emit. Three verdicts: CONSISTENT (≥ 0.85 same-kind) lets it through, DRIFT (0.7–0.85) regenerates, NEW becomes a new persona fact.
Christina M.
User · 87 sessions
User memory
Latest reflection
Welcome back. Last time we left on derivatives. Want to keep going?
Yes, picking up where we left off.
Great. I'll keep the pace gentle since last session ended a bit late.
All six, on every conversation. Live in the dashboard.
Bring your own stack
Vendor-neutral by construction. The provider registry is the only switch — adding Sarvam, Azure, or any new model is a config addition, not an architectural change. Update one line of voice_stack JSON and the next session picks it up. No redeploy, no downtime.
Join the waitlistPick the transcriber that fits your language and budget. Streaming on Deepgram and AssemblyAI; Whisper for batch. Cascading is independent of the LLM and TTS layers, with up to three retries per turn.
Pick the brain your agent actually needs — Anthropic Haiku 4.5 for fast, Sonnet 4.6 for balanced, Gemini for million-token context. If your primary has a bad minute, the next one in line takes the turn and the failover lands in the error_event log.
Pick the voice that holds the persona. Cartesia sonic for speed, ElevenLabs turbo for naturalness, OpenAI TTS for the cheap path. Pronunciation dictionaries (IPA / CMU) override the TTS layer per voice when needed.
From the dashboard, or from your code
Two ways to ship. A no-code Persona Builder for solo founders, clinicians, copywriters, and product teams. A TypeScript and Python SDK for engineers who want it inside their app. Same controls, same versioning, same guardrails.
See what Hyponema agents do in productionCommon use cases people ship on Hyponema
Daily check-ins with patients in chronic care or post-op recovery. The agent remembers symptoms, medication, and last week’s setbacks.
Coaches that pick up after a 14-day pause without rebuilding rapport.
Sessions that resume from the last open question, not a blank slate.
Always-on devices in the Friend / Plaud / Limitless / Bee category that stay on-character across long conversations.
See it in motion
Every turn pre-emit-checked. Tool calls audited. Memory updated server-side at the end of the turn.
The Hyponema MCP server exposes the relational memory engine to any MCP-aware client — Claude Desktop, ElevenLabs, Vapi, Retell, custom agents. Seven tools (retrieve_context, session_resume, narrative_arcs, emotional_trajectory, save_observation, forget_user) over stdio or HTTP/SSE. Bring your own LLM key so reflection and consolidation run on your bill.
Reliable infrastructure
Postgres RLS for isolation. AES-256-GCM envelope encryption for provider credentials. Append-only audit log. The same controls run on every plan, including Free.
Tenant isolation lives in Postgres, not in our API. Every table carries tenant_id with a row-level security policy bound to the session — a bug in our app code can't leak rows.
Provider keys are AES-256-GCM envelope-encrypted: a per-credential DEK wrapped by a KMS-managed KEK. Plaintext lives in memory for the seconds a call lasts, then it's gone. Hyponema operators never see them.
Every persona edit, credential rotation, agent deploy, and DSAR lands in an append-only log with actor, IP, diff, and timestamp. JSON export through one endpoint — your tenant, your retention policy.