autonomous · governable · sovereign

Not a chatbot.
A runtime for persistent AI agents.

Most AI agents answer a one-off request. Pumaï goes further: agents that live in the background, watch topics, run scheduled tasks, build their own tools, draw on long-term memory, call external services — and stay under human control.

24/7background autonomy
~50real tools
3memory tiers
0cross-client leakage
— real-world simulation

A fictional small business. An assistant that truly works.

To test Pumaï under real conditions but with no real-client stakes, we handed the mailbox of an imaginary company — Nautilia Systems, a marine-electronics specialist (GPS beacons, telemetry, monitoring of boats and marinas) — to a fully autonomous assistant that runs around the clock, with no human intervention.

It reads and sorts every incoming email, qualifies quote requests and creates prospect records, tracks support tickets and escalates emergencies, watches competitors and online reputation, chases unpaid invoices with tact, and writes its management reports every day. Everything a sales assistant would do — continuously.

All under strict control: no email is ever sent to an unauthorized recipient, and any committing action (quote, sales reply, complaint) goes out as a draft awaiting approval. The assistant never promises a price, a deadline or a discount that isn't in the rules or data it was given.

a proof-of-work page fed in real time by the instance's actual activity — names and addresses are masked, and nothing is staged.

— what it is

A reliable execution runtime for agents that act

Pumaï brings together, in one runtime, building blocks rarely assembled: autonomy, memory, tools, governance and observability. An agent can take a request, search its memory, pull context via RAG, explore relations via Graph RAG, pick tools, produce a plan, execute actions, ask for approval when needed, log what it does, then consolidate its experience. That complete loop is what most agentic frameworks are missing.

partitioned profiles

Partitioned, yet coordinated

Each profile keeps its private memory, tools and channels. Shared on purpose: a document cabinet and the company directory.

long-term memory

Multi-tier

Recent conversation, summarized segments, vector RAG, relational Graph RAG — consolidated nightly.

agentic engine

Orchestrator / experts

An orchestrator plans, ReAct expert agents execute in parallel waves with real tools.

background autonomy

Heartbeat & Pulz

Watch, trigger and pursue missions without waiting for a user prompt.

tool studio

Tools on demand

Design, test, fix and version tools in a sandbox — plus MCP connectors.

governance

Native, not an add-on

Human approval for sensitive actions, append-only audit log, quotas, partitioning — evaluated at every tool call.

the key difference: Pumaï isn't about "making an AI talk" — it's about giving agents that can act a reliable place to run.

— what's new

The platform just leveled up

A wave of major additions makes Pumaï significantly more powerful: a knowledge graph, a tool-building studio, MCP connectors, a phone gateway (inbound/outbound calls), streaming responses with visible reasoning, an interface now in English, and more.

explore the new capabilities →

— the paradigm shift

More than an assistant. A runtime that acts.

Classic assistants wait, forget, and stop at text. Pumaï addresses all three limitations at once.

axis classic assistant pumaï
Trigger Passive — waits for a prompt Proactive — a background heartbeat, lives without human interaction
Memory Amnesic — limited context window Sedimentary — partitioned memory + nightly consolidation
Action Simulates, or stops at text Agentic — an orchestrator delegates to a team of expert agents with real tools
Governance Uncontrollable black box Glass box — conditional locking, human validation, append-only audit

put another way: Pumaï is less an "AI assistant" than an operating system for AI agents.

— architecture

The architecture, at a glance

A central API resolves the profile, assembles context, applies security and quota — then routes into three subsystems: memory, the agentic engine, and the autonomous background loop.

Web
REST API
Telegram
Discord
Phone
Central API · FastAPI profile resolution · context assembly · security · token quota
on every request ↓
memory
Narrative memorypersona · voice · durable facts
PostgreSQL + pgvectorsegments · RAG (HNSW 1024-d) · Graph RAG
agentic engine
Orchestratorplans, never executes
Expert agentsReAct loops · ~50 tools · MCP
Lock Managerarbitrates every sensitive call
background loop
Heartbeatticks every N min → Pulz
Schedulernightly dreams · self-analysis
Outbound channels email (SMTP) · Telegram · web · phone — every outbound action passes through the Lock Manager
system map — channels → API → memory · action · autonomy → outputs

stack — Python 3.12 · FastAPI · PostgreSQL + pgvector · OpenAI-compatible client (model provider configurable) · Docker.

— sovereign & tailored

Your assistant, your rules, your data

Pumaï is deployed as an isolated instance: each client gets their own — their data, their mailbox, their keys, their configuration — isolated from other clients'. You stay in control end to end.

partitioned data

One instance per client

Your data is stored in a dedicated instance, isolated from other clients'. Any processing or transfer to external services (AI model, email, telephony…) depends on the providers and connectors you configure.

model of your choice

No vendor lock-in

Pumaï is tied to no AI provider: we pick the model that fits your requirements for quality, cost or confidentiality — and the system reasons in any language.

extensible

Your tools, your connectors

The list of capabilities isn't fixed: custom tools built in a sandbox, MCP connectors to your services (mail, calendar, databases…). We open a tool, the assistant uses it.

traceable

Never a black box

Every decision, every action and its cost is logged and consultable: append-only audit log, execution trace, per-instance quotas. An AI that acts on its own, yet always under your watch.

— new capabilities

What just landed in Pumaï

01 — memory

Knowledge graph

Beyond vector RAG, memory now weaves a queryable graph of entities and relations — better-connected answers and an understanding that structures itself over time.

02 — action

Tool-building Studio

Design, test and version your own tools through a guided, sandboxed assistant. The agent gains new capabilities, tailored to your needs.

03 — action

MCP connectors

Plug in external MCP servers to extend the agent: add manually, search the npm registry, or import directly from a URL, package or JSON.

04 — control

Redesigned Tools page

A reworked, far more practical interface, with a Lock Manager that finally makes approving and locking each tool simple.

05 — channels

Phone gateway

Through Twilio, your agent can now receive and place phone calls — a voice that answers, calls out and reports back.

06 — experience

Streaming responses

Messages appear progressively as the AI writes them — smooth on screen, essential for real-time voice.

07 — transparency

Execution trace

Follow the agent's execution trace right in the chat — its steps and the rationale behind its decisions, not just the final answer.

08 — observability

More powerful logs

The log page becomes a real activity view: sessions, runs and events, to understand what the system is doing.

09 — international

English interface

The interface is now translated into English (≈80%, the rest is on its way), opening Pumaï to a wider community.

10 — support

Built-in support

A support button at the bottom of the interface sends a message that lands straight in our tracking tool.

11 — competences

Reusable Competences

Teach a know-how once — handling a quote, managing a complaint, opening a ticket — and the assistant then applies it identically. A dedicated studio builds them through a simple guided interview and measures the gain.

12 — action

A real mailbox

The assistant reads, files, opens attachments, prepares drafts and actually sends its replies — but never to a recipient outside your allowlist, with the "draft awaiting approval" option.

13 — memory

The document cabinet

A shared central binder: you drop in your reference documents, the assistant files its output (reports, quotes, sent attachments). Everything is dated, traced, and its lifecycle tracked (draft → to send → sent).

14 — transparency

Activity dashboard

Not a technical log but a narrative: goal → steps → result. Ongoing missions, the day's figures, time and cost saved — auditable and honest (a lookup is never presented as an action).

15 — proof

Showcase — proof of work

A page that shows, live, from the system's real log, everything the assistant accomplishes autonomously — honest figures, decisions, downloadable files. Private by default, publishable in one click, personal data masked.

— in development

Next step: contexts twice as light

The next big leap is already underway: representing context no longer as a long text to re-read, but as a knowledge graph seasoned with prompt fragments — a far more compact form that cuts 50 to 70% of the volume the assistant must load at each step, without losing the thread. The payoff: faster answers, a markedly lower cost per action, and a longer working memory at equal budget.

This is still research & development: the 50% compression mark is already reached in testing, with near-intact comprehension. The current work targets ~60% while keeping comprehension nearly intact. We advance, step by step.

prototype in testing — not yet in production.

— the glass box

An AI that acts on its own must stay controllable

Pumaï doesn't promise a magical, uncontrollable AI: it offers infrastructure for governable agents. Every action can be traced. Every tool can be limited. Every profile can be isolated. Every sensitive decision can go through human approval. And all autonomy can be disabled instantly — with a single control.

governance in detail →

— go further

Let's talk about your project

Have a specific use case to scope, or want to see Pumaï run on your business? Write to us — individual reply within 48h.

No mailing list — just a real conversation.

individual reply — no mailing list

Your data (first name, email, message) is used only to answer your request, by the data controller (Polymathe SI). See the privacy policy.

message received — reply within 48h
send error — please retry