Contextual memory that lets your agent Remember.
Mnemix gives AI voice agents persistent caller memory and real-world enrichment. One API call before every interaction — your agent knows who's on the line, what happened last time, and who they are. Designed for sub-300ms voice recall.
Voice is where it starts. Underneath is contextual intelligence — a memory substrate any agent can build on.
From zero to enriched recall in two commands.
$ npm i -g @mnemix-ai/cli $ mnemix recall +15551234567 → Mike Reynolds · 3 calls · fleet, commercial "Repeat customer. Fleet account. Prefers mornings." enrichment: carrier · line type · company
Most AI agents have amnesia.
Every call, every chat, every session starts cold. The agent re-introduces itself, re-asks the same questions, and forgets the relationship the moment the line drops. Your customers notice. They've talked to you five times — and your “AI” acts like it's the first.
Ring → recall → enrich → answer.
The moment a call connects, your agent calls recall_and_enrich. Memory comes back instantly from cache; enrichment fills in asynchronously in the background so the voice path never blocks. Designed for sub-300ms voice recall.
One round trip. The whole context.
Call it the instant the phone rings and it returns — in a single request — who the caller is, your history with them, and live enrichment from the real world. Your agent opens its mouth already knowing the context a human would have.
timing_ms.total is a real response field — never a hardcoded benchmark.
// POST /v1/recall_and_enrich
{
"phone_number": "+15551234567",
"trigger": "answered",
"session_id": "call_abc123"
}
// → 200 · timing_ms.total: 286
{
"known": true,
"caller": {
"name": "Mike Reynolds",
"total_calls": 3,
"tags": ["fleet", "commercial"]
},
"memory": {
"summary": "Repeat customer, 3 calls over 2 months. Fleet account. Prefers mornings.",
"recent_calls": [
{ "summary": "Confirmed brake inspection for Thursday", "outcome": "confirmed" }
]
},
"enrichment": {
"person": { /* … */ },
"company": { /* … */ },
"phone": { /* … */ }
}
}Deterministic by design. Auditable by default.
Mnemix is bi-temporal: it remembers not just what's true, but what was true and when you learned it. Every fact is versioned across two timelines — when it happened in the world, and when your system came to know it.
That makes your agent's decisions reconstructable. Given the same context as of any moment in the past, it behaves the same way — every time. You can replay exactly what the agent knew when it made a call, and prove the decision against the state-of-record as of that instant.
It's the difference between an agent that guesses and one you can audit— whether it's confirming an appointment or reconciling a ledger.
Valid-time (when it was true) and transaction-time (when you knew it). Corrections never overwrite; they version.
Reconstruct the exact context the agent saw at any past moment.
The substrate guarantees a single, ordered version of every fact. Defensible, not approximate.
Everything a voice agent needs to remember.
Sub-300ms voice recall
Built for the hot path, not a batch job. Designed for sub-300ms voice recall.
Persistent memory
Every call summarized and remembered; relationships compound instead of resetting.
Real-world enrichment
Caller identity, carrier, and line type via Trestle, Twilio Lookup, and Baylio — one clean SDK shape.
Bi-temporal & auditable
As-of reconstruction; deterministic, defensible decisions you can replay.
Tenant-isolated
Every tenant's data is isolated at the database layer, not just the app layer. Your callers are yours.
MCP-native
Drop Mnemix into any MCP-compatible agent with built-in tools for lookup, save, search, and enrich.
Why not a general-purpose memory library?
General memory tools are built for chat: store text, embed it, retrieve later. Voice is a different animal.
General-purpose memory
- Built for chat: store text, embed, retrieve later
- Keyed on a paragraph of past messages
- Seconds of retrieval budget — fine for async
- No real-world identity; just a vector of history
- Approximate recall, no audit trail
Mnemix — voice-native
- Purpose-built for the voice hot path
- Phone-keyed — a number, not a paragraph
- Designed for sub-300ms voice recall
- Real-world enrichment: identity, carrier, line type
- Bi-temporal — context is auditable, not approximate
Start free.
- Building, testing, side projects
- Free Hobby key — no card
- Community Discord
- Production voice apps at volume
- Higher limits + enrichment spend
- Priority support + SLA
Questions, answered.
What does Mnemix actually return?+
A caller's profile, your recent-call memory, and real-world enrichment — in one call, designed for sub-300ms voice recall.
What powers the enrichment?+
Trestle, Twilio Lookup, and Baylio, returned in one normalized shape (person / company / phone).
Is my caller data isolated?+
Yes — tenant isolation is enforced at the database layer, not just in application code.
What's “bi-temporal”?+
Mnemix versions every fact across two timelines, so you can reconstruct exactly what your agent knew at any past moment. Deterministic, auditable.
Do I need to use HTTP?+
No — there's a native MCP server, so MCP-compatible agents get memory tools directly.
Is it only for voice?+
Voice is the wedge; the substrate underneath is general-purpose contextual intelligence — a memory layer any agent can build on.
Give your agent a memory.
Designed for sub-300ms voice recall. Free to start. We're onboarding beta testers now — drop your email and we'll send a Hobby key as spots open.
Early access for builders shipping AI voice agents. Use your work email for better early-access routing. No spam, no fake scarcity.
Submits your email to the Mnemix early-access waitlist. Use your work email for better early-access routing.
Last updated: .