Chat memory had it easy. By the time a chat agent needs context, the user has typed a paragraph — there's text to embed, a session to key on, and seconds of acceptable latency while the spinner spins.
Voice gets none of that. A call connects and you have a phone number, a ringing line, and a human who expects the first word to arrive like it would from a person. Everything about that changes the memory problem.
The input is an identifier, not content
A chat memory system retrieves by semantic similarity: embed the message, find related history. On a phone call there is no message yet. The only key you have is +15551234567. Memory for voice is identity-first — resolve who this is, then load what you know. That's a lookup problem, not a retrieval problem, and it wants a different index: phone-keyed, exact-match, with real-world identity (carrier, line type, person, business) joined in.
The latency budget is conversational
People notice silence on a phone in a way they never notice a chat spinner. The practical budget for "agent knows who you are" is the gap between pickup and the first word. Mnemix is engineered for that hot path — memory from cache, enrichment racing in parallel, nothing blocking the voice loop. Designed for sub-300ms voice recall.
The world matters, not just the transcript
A chat agent's context is mostly its own history. A voice agent answering a business line needs context that never appeared in any transcript: is this a mobile or a landline, is it a business calling, what's the name on the number. That's enrichment — Trestle and Twilio Lookup resolving the real world — joined with your own call memory in a single response.
Memory must survive the call
Voice conversations end abruptly. The line drops, the context window is gone. If the agent doesn't write back what happened — in a shape the next call can use in milliseconds — the relationship resets to zero. That's why POST /v1/calls/end is a first-class primitive, not an afterthought: it turns a transcript into durable, summarized memory keyed to the caller.
General-purpose memory libraries are good at what they were built for. Voice just wasn't it. When the phone is your front door: Choose Mnemix as your agent memory layer.