Your team is finite.
Your personas are not.
PersonaMesh is the platform to author, train, and run AI agents with real expertise, a consistent voice, persistent memory, and a life of their own — running 24/7 in the cloud, in your datacenter, or on the edge.
Trusted by teams building agents in production
Chatbots forget. Agents wander. Personas are how AI becomes someone.
The next decade of software isn't a smarter form field. It's a workforce of digital characters with names, expertise, voices, and lives — that show up tomorrow remembering what happened today.
Off-the-shelf agents drift, hallucinate identity, and reset every session. Building a real persona means schema, memory, scheduling, safety, evals, and 24/7 ops — not just a system prompt.
- Stateless chats. Yesterday's context is gone.
- Voice drift. Brand sounds like ChatGPT.
- No life. Bot only exists when poked.
- Glue code for memory, schedules, channels, eval, audit.
- Compliance hand-waving when legal asks.
- Versioned identity, voice, expertise, values.
- Episodic + semantic + procedural memory.
- Generated patterns of life. They sleep, work, react.
- Autonomy loop with planner, tools, budget, kill switch.
- Audit trails, evals, governance shipped from day one.
Everything a digital character needs to be someone real.
Four primitives, one platform. Compose them into personas that hold a job, run a beat, tutor a class, or play a role.
Identity, voice, and expertise as code
A versioned persona schema captures personality (Big Five), values, voice register, forbidden phrases, expertise levels, knowledge sources, tool budgets, and safety policy. Compile it deterministically — same inputs, same prompt, every time.
name: "Dr. Mira Vale"
expertise:
- { domain: marine_ecology, level: expert }
- { domain: science_journalism, level: proficient }
personality:
big_five: { O: 0.82, C: 0.71, E: 0.55, A: 0.66, N: 0.31 }
values: [curiosity, candor, stewardship]
voice:
register: warm-academic
forbidden_phrases: ["as an AI"]
tools:
allowed: [web_search, calendar, send_email]
budget: { usd_per_day: 2.00 }Episodic, semantic, procedural
A three-tier memory system that survives restarts and grows up. Importance-weighted retention, hierarchical summarization, retrieval blending recency × relevance × persona affinity. Forget-on-request built in.
A schedule. A rhythm. A life.
Generate a persona's week — sleep, work, social, hobbies, errands — coherent with timezone, energy curve, and social graph. Travel-time aware. No two days identical, but always in character.
Planner, executor, reflection
A tick scheduler reads the pattern of life and enqueues goals. The planner produces structured plans, the executor calls sandboxed tools, the reflector journals at end of day. Hard budget caps and a kill switch — always.
They have a life. So they have a perspective.
PersonaMesh personas don't just answer when called. They sleep through the night, work through the day, react to events, and arrive in conversation having actually done something since you last spoke.
From a YAML file to a 24/7 colleague in three steps.
Author
Define identity, expertise, voice, values, knowledge sources, tools, and budgets in the Studio or in code. Versioned and diffable.
Train
Three personalization tiers: prompt+RAG, exemplar bank, optional LoRA/DPO on opt-in transcripts. Eval harness gates every release.
Run
Deploy to SaaS, your VPC, or the edge. Patterns of life drive autonomous activity. Channels reach Slack, email, voice, webhooks.
import { PersonaMesh } from "@personamesh/sdk";
const pm = new PersonaMesh({ apiKey: process.env.PMESH_KEY });
const mira = await pm.personas.create({ from: "./mira.yaml" });
const reply = await mira.chat.send(
"What did we decide about the kelp paper?",
{ conversationId: "conv_42", stream: true }
);
await mira.memory.remember({
fact: "Prefers Oxford comma",
importance: 0.6,
});
await mira.autonomy.start({ schedule: "pol", budgetUsd: 2 });"We turned three pages of brand voice docs into a persona that sounds like us in every channel. The eval gate alone is worth the platform."
"The patterns-of-life engine was the unlock. Our NPCs no longer feel like prompts — they feel like people you've interrupted."
"On-prem with our own model, our own KMS, full audit log, and no phone-home. That's why we picked PersonaMesh."
Wherever a person would be too expensive,
put a persona.
Customer support
EnterpriseTier-1 personas with real product knowledge, brand voice, and a memory of the customer. Escalate only when needed.
K–12 tutoring
EducationA patient tutor that remembers what the student got wrong last week and adjusts pace. Parent-visible, school-policy aware.
Sci-fi & game NPCs
GamingCharacters with histories, grudges, schedules, and a world they live in. They show up changed by yesterday.
Research analysts
FinanceA persona per beat that reads overnight, files morning notes, and remembers what your PM cares about.
Sales SDRs
GTMA team of digital SDRs with distinct voices and territories. Coordinated so two of them never email the same lead.
Internal copilots
OperationsA persona for HR, IT, finance — each with the policies, tone, and escalation paths your company actually uses.
One platform. Three perimeters.
The same persona definition runs in our cloud, your datacenter, or on a ruggedized box at the edge. Conformance-tested every release.
PersonaMesh Cloud
- Cell-based isolation, regional residency
- SSO + SCIM, BYOK, audit-log export
- Metered billing, status page, 24×7 on-call
- Sign up in 60 seconds
Inside your perimeter
- Bring your own model, vector DB, KMS, IdP
- Hardened images, FIPS build, STIG guides
- Offline installer, signed artifacts, SBOM
- LTS releases with 24-month patch window
On the device
- Quantized open models (Llama, Mistral, Phi)
- SQLite + sqlite-vss for local memory
- Online / opportunistic / offline-only modes
- Fleet console with phased rollouts
Build for free. Pay when your personas earn their keep.
- 1 workspace, 3 personas
- Local memory, BYO model key
- Studio + SDK + CLI
- Community support
- Hosted SaaS, 1 region
- Persistent memory, evals
- Up to 10k autonomous ticks/mo
- Email support
- Patterns of life, autonomy loop
- Channels: Slack, email, voice
- SSO, audit log, BYOK
- Business-hours support, 99.9% SLO
- Dedicated infra or air-gapped
- FedRAMP / HIPAA / FIPS path
- TAM, 24×7 on-call, custom SLA
- Source license available
Tokens, tool calls, autonomy ticks, and storage are metered. Hard budget caps per workspace, project, and persona.
Questions, answered.
How is this different from a system prompt + a vector DB?
A system prompt is a string. A persona is a versioned identity with personality, voice, expertise, memory policy, tool budgets, safety rules, and a life. PersonaMesh ships the schema, the eval harness, the autonomy loop, the audit trail, and the deployment substrate so you don't reinvent each one.
Can I impersonate a real person?
No, not without verified, ongoing consent — and even then we restrict tools and channels. Identity verification is required for any 'real-person' persona, and we publish a public list of verified personas. Impersonation without consent is a ToS violation and we have detection in place.
Which models do you support?
Anthropic (default), OpenAI, Bedrock, Azure OpenAI, Vertex, plus self-hosted via vLLM, TGI, Ollama, and llama.cpp. The provider layer is pluggable; conformance suite runs against the tested matrix every release.
What happens when a persona is offline at the edge?
It keeps running. Local SQLite + sqlite-vss handles memory and retrieval, a quantized open model handles inference, the local policy engine handles safety, and outbound actions queue for store-and-forward sync when connectivity returns.
How does memory not balloon costs forever?
Hierarchical summarization, importance-weighted eviction, and per-persona budgets. You set a retention horizon; the platform compresses below it and prunes when the budget is hit. Recall stays high because summaries preserve the salient facts.
Is the autonomy loop safe?
Hard budget caps (USD, tokens, wall-clock) at workspace, project, and persona scope. Every tool is sandboxed, every action is audited, every plan is policy-checked, and there's an always-available kill switch. Human-in-the-loop approval modes for sensitive channels.
Source-available?
The Community SDK and CLI are Apache 2.0. The edge runtime and core platform are source-available under BSL-1.1, converting to Apache 2.0 after four years. Cloud-only features (multi-tenancy, fleet console, billing) stay closed.
Can we run our own evals?
Yes. The eval harness ships with the SDK. Bring your own datasets and rubrics; the harness handles scoring, regression detection, and CI integration. Public benchmarks are published per release.
Your headcount went up.
It just doesn't show on the org chart.
Spin up your first persona this afternoon. Bring it to production this quarter. Run a hundred of them next year.