agntspace 1.38.0

Personal AI agent that runs your life alongside you

AgntSpace

Personal AI operating system / second brain. Run it in the cloud at agntspace.com, or locally on your machine.

pip install agntspace && agntspace start
# → opens http://localhost:8000, onboarding wizard guides you through setup

New here? Read the 5-minute tour. Extending AgntSpace? See the community hub for skills, plugins, and vault templates.

Origin Story

AgntSpace started as a personal AI assistant with a memory system — a way to give an LLM persistent context about your life, projects, and decisions across conversations. We built layered memory compaction, a knowledge graph, trust evolution, and contract enforcement. It worked, but the knowledge side felt incomplete.

The missing piece was an LLM-powered wiki — a system where an AI reads your raw sources and compiles them into an interconnected, always-up-to-date knowledge base with wiki-links, entity pages, and source provenance. We wired that concept into the agent architecture: the agent ingests your files, compiles wiki articles, cross-links entities, detects contradictions, tracks provenance back to sources, and keeps the whole thing current as new information arrives. Combined with the memory system, you get an AI that not only remembers what you told it — it understands what you know.

Philosophy

Your AI agent will learn how you think, how you work, and how you decide. That makes the persistence layer — your memory, your knowledge graph, your behavioral context — the most valuable digital asset you own. It should not belong to the platform that hosts the model. It should not leave with your employer when you change jobs. It should not be locked in a proprietary format that makes switching providers an act of cognitive amnesia.

AgntSpace exists because we believe the persistence layer belongs to you: stored in plain files you can read without us, governed by a contract you write and the agent enforces in code, with a complete decision trail so "what did you do?" always has a verifiable answer. The agent earns autonomy through demonstrated competence. It never acts outside the contract. And when you leave, your second brain comes with you — no export, no migration, no permission needed. It was never ours.

Every architectural choice in this project flows from that commitment:

• The Contract is how "governed by an agreement you write" becomes code — YAML permissions, default-deny, enforced on every action.
• The Trust System is how "earns autonomy through demonstrated competence" works — feedback-driven, domain-specific, reversible.
• The Decision Log is how "complete trail of every decision" is kept — correlation IDs, memory provenance, per-tool-call accountability.
• The Vault is how "belongs to you" is guaranteed — plain markdown, no proprietary format, works with Obsidian or any text editor, no cloud required.
• The Epistemic Engine is how "never pretends" is enforced — five-factor knowledge weights, justification chains, contradiction detection, verified vs. believed distinction.

What Is AgntSpace?

AgntSpace is a personal AI agent that runs your life alongside you. Not a chatbot you visit when you have a question — a persistent partner that knows your projects, tracks your goals, manages your communications, learns from your feedback, and grows more capable over time.

You talk to a main agent through whatever channel you prefer — web, Telegram, WhatsApp, Discord, Slack, CLI. Behind it, a team of specialized subagents works on your behalf: a Researcher digs into topics, a Writer drafts in your voice, an Analyst crunches numbers, a Planner breaks goals into steps. You can deploy as many subagents as you need — pre-built or custom — and the main agent orchestrates them all under your contract.

The agents read and write your knowledge in a markdown vault you can edit with any tool you like. They delegate your work, coach your goals, and consolidate what you learned today into long-term memory while you sleep.

Run it locally and everything stays on your machine. Or use agntspace.com and we host it for you. Either way: your data is stored in markdown files, your agent's personality is readable text, and the rules of engagement between you are an editable contract — no black boxes.

Strategic Independence

AgntSpace is designed to be fully provider-independent as a core architectural principle:

• LLM independent. Switch between Anthropic, OpenAI, Google, Mistral, Ollama, or any OpenAI-compatible API at any time. Your data, contract, memory, and trust scores stay the same. The LLM is a replaceable reasoning engine, not a platform lock-in.
• Infrastructure independent. Run on your laptop, a server, Docker, or our hosted service. Move between them freely. Your vault is portable markdown — not tied to any cloud, database, or vendor.
• Channel independent. Add or remove messaging channels (Telegram, WhatsApp, Discord, Slack) without affecting the agent's capabilities. Channels are delivery mechanisms, not dependencies.
• Storage independent. Integrate with Obsidian, Logseq, or any markdown-based tool. Your vault works with or without external editors.

This means you are never locked into a specific AI provider, cloud platform, or tool ecosystem. AgntSpace is yours to run however you want, with whichever providers you choose, forever.

graph TB
    subgraph Channels["How You Connect"]
        WebUI["Web UI"]
        Msg["Telegram / WhatsApp\nDiscord / Slack"]
        Desktop["Desktop App"]
        CLI["CLI"]
    end

    subgraph Core["AgntSpace"]
        Main["Main Agent"]
        Sub["Subagents\nResearcher / Writer\nAnalyst / Planner"]
        Main -->|orchestrates| Sub
    end

    subgraph Brain["The Brain"]
        Vault["Vault (MD)"]
        Skills["Skills (MD)"]
        Memory["Memory (SQLite)"]
        Domains["Domains\n(Life OS)"]
    end

    subgraph LLM["LLM Provider"]
        Cloud["Anthropic / OpenAI"]
        Local["Ollama (local)"]
    end

    subgraph Integrations["Integrations"]
        Gmail["Gmail"]
        GCal["Google Calendar"]
        GitHub["GitHub"]
    end

    WebUI & Msg & Desktop & CLI --> Main
    Main --> Vault & Skills & Memory & Domains
    Main --> Cloud & Local
    Main --> Integrations

---

What Makes This Different

Most AI products give you a conversation. AgntSpace gives you a relationship — with identity, trust, memory, growth, and accountability.

1. The Contract — A Negotiated Relationship

Every other AI system has terms of service written by the company. AgntSpace has a contract written by you.

CONTRACT.md is a human-readable agreement between you and your agent that defines: what it can do, what it can't, how autonomous it should be, what data it can send to the cloud, and when it should ask before acting. The contract is not just documentation — it's enforced in code. Every agent action is checked against your contract before execution. When the contract is ambiguous, the agent asks. It does not assume permission.

No other personal AI system formalizes the human-agent relationship as an editable, enforceable agreement.

2. Three Identities, One Relationship

AgntSpace has three entities, not two:

• You — your identity, preferences, and context live in USER.md and PROFILE.md
• The agent — its personality, values, and boundaries live in SOUL.md
• The agreement — the rules of engagement live in CONTRACT.md

You can open SOUL.md and read exactly who your agent is. You can edit it. You can change its personality to match how you want to be supported. No other AI product gives you a readable, editable file that defines the agent's identity.

graph LR
    Human["You\nUSER.md + PROFILE.md"]
    Contract["CONTRACT.md\n(enforced in code)"]
    Agent["Agent\nSOUL.md"]

    Human <-->|"negotiated\nagreement"| Contract
    Contract <-->|"governs\nbehavior"| Agent
    Human <-->|"relationship"| Agent

    style Human fill:#4A90D9,color:#fff
    style Contract fill:#F39C12,color:#fff
    style Agent fill:#27AE60,color:#fff

3. Trust That Evolves

Other AI systems have a fixed level of capability. AgntSpace has trust that grows from experience.

The agent starts cautious. As you delegate tasks and give feedback, it earns trust per domain. After 5 successful email drafts, it earns autonomy over email. If it sends something wrong, trust drops. You can override trust at any time, but you shouldn't have to — it adapts.

This means a new AgntSpace agent and a 6-month-old AgntSpace agent behave differently. The one that's been working with you is more capable, more autonomous, and more aligned with how you work — because it earned that through demonstrated competence, not a settings toggle.

4. The Agent Teaches Itself

When the agent notices it keeps doing the same kind of task — investor update emails, meeting notes, financial reports — it drafts a skill for it. Not code. A structured instruction set (SKILL.md) that captures the pattern, your preferences, and the right approach.

You review the skill. If it's good, you approve it. Now the agent uses it going forward — and continues refining it as more feedback comes in.

No other system has the agent composing and refining its own instruction sets from observed patterns and accumulated feedback.

5. A Shared Thinking Space

The daily journal isn't a log — it's a shared cognitive space where both you and the agent think.

You write your thoughts, reflections, and decisions. The agent writes its observations, actions, and learnings. Both are clearly attributed. At end of day, the reflection process consolidates what matters into long-term memory.

This means you can open today's journal in any markdown editor and see — in one file — what you thought, what the agent noticed, what decisions were made, and what was learned. No other system gives both human and agent a shared, transparent place to think.

6. Knowledge Graph — Automatic

Other AI tools search by keywords. AgntSpace builds a knowledge graph from every interaction — automatically.

The agent creates [[wiki-links]] in memory files as it works. Every person, project, decision, and concept gets connected. When you ask "What do I know about the Acme deal?" — the agent doesn't just search text. It traverses the graph: Acme Corp → Sarah Chen → B2B Partnership → Q2 Pivot → Investor Update. You get the full picture, not just keyword matches.

You never create links manually. You never manage a graph. The agent builds it from every conversation, email, task, and decision. The graph view in the UI is a visualization of what the agent already knows.

7. A Coach, Not Just a Tool

AgntSpace doesn't just execute tasks — it tracks your goals, predicts your behavior patterns, and gently nudges you when things stall.

"You tend to lose momentum on side projects after 2 weeks. The book outline hasn't been touched in 10 days."

It knows this because it's been reading your journal, tracking your tasks, and consolidating your patterns through weekly reflection. It's not a generic productivity tip — it's based on your actual behavior over time.

7. Honest Privacy

Every "private AI" product claims your data is safe. AgntSpace tells you exactly what happens:

• Your vault is local. Your embeddings are computed locally. Your search is local.
• But when the agent thinks, it sends context to a cloud LLM. Your notes, your profile, your memory — transmitted to Anthropic or OpenAI.
• Sensitive directories (finances, health, journal, goals) require your confirmation before sending.
• For full privacy, switch to Ollama. All reasoning stays on your machine. The architecture supports this as a first-class mode, not an afterthought.

No disclaimers buried in settings. Privacy is a first-class design principle.

8. Multi-Agent Orchestration

The product is called AgntSpace because it's not one agent — it's a space where agents work together under your contract, orchestrated by a main agent you trust. You build the team you need.

graph TD
    Human["You: 'Prepare the Q2 board deck'"]
    --> Main["Main Agent\n(decomposes task)"]
    Main --> R["Researcher\npull Q2 metrics"]
    Main --> A["Analyst\ncreate charts"]
    Main --> W["Writer\ndraft summary"]
    R --> Main2["Main Agent\n(assembles result)"]
    A --> Main2
    W --> Main2
    Main2 --> Human2["Delivers deck → You review → Feedback"]

    style Main fill:#4A90D9,color:#fff
    style Main2 fill:#4A90D9,color:#fff
    style R fill:#27AE60,color:#fff
    style A fill:#E67E22,color:#fff
    style W fill:#8E44AD,color:#fff

9. Everything Is a File

Your identity. Your agent's soul. The contract between you. Your memory. Your tasks. Your trust scores. Your goals. Your skills. All of it is markdown files you can open in any text editor.

There is no proprietary database that holds your life. There is no cloud account that locks you in. If AgntSpace disappears tomorrow, you still have every file, in a format that works with any markdown editor — Obsidian, VS Code, Logseq, Notepad, or whatever comes next.

The database (SQLite) is a search index. Delete it and it rebuilds from your files. Your files are the truth.

---

The Cognitive Loop

graph LR
    Identity["Identity\nSOUL + USER + CONTRACT"]
    --> Thinking["Thinking\nJournal"]
    --> Delegating["Delegating\nSubagents"]
    --> Doing["Doing\nTasks"]
    --> Remembering["Remembering\nReflection → MEMORY"]
    --> Learning["Learning\nFeedback → Trust"]
    --> Growing["Growing\nSkillSchool"]
    --> Coaching["Coaching\nGoals + Nudges"]
    --> Autonomy["Autonomy\nTrust evolves"]
    --> Identity

    style Identity fill:#4A90D9,color:#fff
    style Thinking fill:#7B68EE,color:#fff
    style Delegating fill:#E67E22,color:#fff
    style Doing fill:#27AE60,color:#fff
    style Remembering fill:#8E44AD,color:#fff
    style Learning fill:#F39C12,color:#fff
    style Growing fill:#1ABC9C,color:#fff
    style Coaching fill:#E74C3C,color:#fff
    style Autonomy fill:#3498DB,color:#fff

---

Requirements

Depends on how you run it:

	agntspace.com	Desktop App	CLI
You need	A browser	Download installer	Python 3.11+
OS	Any	Windows, macOS, Linux	Windows, macOS, Linux
LLM	Included (cloud)	Cloud: API key. Local: Ollama installed	Same as desktop
Hardware (cloud LLM)	Nothing	2GB RAM, any CPU	Same
Hardware (local LLM)	N/A	16GB+ RAM, modern CPU/GPU	Same

That's it. No Docker required. No database to install. No Redis, no Postgres, no Node.js. SQLite is bundled with Python. FastEmbed downloads its model on first run (~80MB).

LLM — The Brain Behind the Agent

Mode	Provider	What you need	Quality	Privacy	Cost
Cloud	Anthropic (Claude)	API key from console.anthropic.com	Best	Vault content sent to provider	Pay per use (~$3-15/mo typical)
Cloud	OpenAI (GPT)	API key from platform.openai.com	Great	Vault content sent to provider	Pay per use
Local	Ollama	Install Ollama + download a model	Good (depends on model)	Total privacy — nothing leaves your machine	Free (your hardware)

---

Getting Started

Option A: agntspace.com (nothing to install)

Sign up at agntspace.com and start talking to your agent in the browser. Same features, hosted for you. Export anytime.

Option B: Desktop App (no terminal)

Download for Windows | macOS | Linux — double-click, visual setup wizard, runs in system tray.

Option C: One-Line Install

macOS / Linux:

curl -fsSL https://agntspace.com/install.sh | bash

Windows (PowerShell):

irm https://agntspace.com/install.ps1 | iex

Or manually:

pip install agntspace
agntspace init

Interactive setup: creates ~/.agntspace/, configures LLM (cloud or Ollama), runs onboarding conversation to generate your PROFILE.md and CONTRACT.md.

agntspace start              # Start server
agntspace chat               # Talk via CLI
agntspace delegate "task"    # Delegate directly
agntspace journal "thought"  # Quick journal entry
# Web UI → http://localhost:8000

No Markdown Required

Everything works through the web UI. Power users can edit vault files directly — changes sync instantly.

Action	Web UI	File (optional)
Your profile	Control Panel → Profile	vault/PROFILE.md
Agent permissions	Control Panel → Contract	vault/CONTRACT.md
Agent personality	Control Panel → Agent	vault/SOUL.md
Journal	Journal page	vault/journal/*.md
Tasks	Tasks board	vault/tasks/*.md
Goals	Goals page	vault/goals/*.md
Subagents	Control Panel → Agents	vault/agents/*.md
Trust scores	Control Panel → Trust	vault/TRUST.md

---

Key Concepts

Task Delegation

Mode	Behavior
Just do it	Agent executes, reports when done
Review	Agent executes, you approve before finalizing
Recommend	Agent researches, presents options, you decide
Watch	Agent monitors, alerts when triggered

Tasks are markdown files with progress logs, escalation notes, and feedback. Agent escalates gracefully when stuck.

graph LR
    D["Delegated"] --> P["In Progress"]
    P --> B["Blocked\n(escalate)"]
    B --> P
    P --> R["Review"]
    P --> Done["Done"]
    R --> Done
    R --> Rev["Revision\nNeeded"]
    Rev --> P
    Done --> F["Feedback\ngood / redo"]
    F --> T["Trust\nupdated"]

    style D fill:#3498DB,color:#fff
    style P fill:#F39C12,color:#fff
    style B fill:#E74C3C,color:#fff
    style R fill:#8E44AD,color:#fff
    style Done fill:#27AE60,color:#fff
    style F fill:#1ABC9C,color:#fff
    style T fill:#4A90D9,color:#fff

SkillSchool

Agent notices repeating patterns, drafts a skill (SKILL.md), proposes it for your review. Three triggers: pattern detection, feedback accumulation, human request. Skills refine over time as feedback continues.

Journal (Shared Thinking Space)

Daily file, four sections: Thoughts (human), Observations (agent), Decisions (both, with reasoning), Learnings (both). Thinking-out-loud detection in chat: agent asks "Save to journal?" Configurable: ask, auto, or off.

Reflection

End-of-day: journal → reflection draft → (your approval or auto) → MEMORY.md updated. Weekly: 7 days synthesized into patterns and progress. Feeds coach, SkillSchool, and trust recalculation.

Coach

Tracks goals (vault/goals/*.md), predicts behavior patterns, nudges gently on stalls, captures aspirations from conversation, celebrates wins. Opt-in, configurable frequency and quiet hours.

4-Tier Proactivity

Tier	Behavior
Observer	Only responds when asked
Advisor	Suggests but never acts
Assistant	Acts on routine, asks for non-routine
Partner	Full autonomy within contract boundaries

Autopilot Mode

Enable for travel/busy periods with expiry date. Reflections auto-commit, trusted domains auto-approve, notifications batched to daily digest. Auto-disables with welcome-back summary.

Honest Privacy

Always local: vault, embeddings (FastEmbed), search index (SQLite), heartbeat state. Sent to cloud LLM (cloud mode): SOUL.md, USER.md, MEMORY.md excerpts, vault content for current task, conversation history. Sensitive dirs (finances, health, journal, goals) gated — agent asks first. Full privacy: switch to Ollama. All reasoning on-device.

Everything Is a File

Identity, memory, tasks, trust, goals, skills, subagent definitions — all markdown. Database is a derived search index. Delete it, it rebuilds. No proprietary format. No vendor lock-in.

---

Proactive Systems

System	Frequency	What it does
Heartbeat	Every 30 min	Gathers integration state, diffs, logs to journal, notifies on changes
Task scheduler	Continuous	Checks pending/watch tasks, triggers on conditions
Daily reflection	End of day	Journal → reflection draft → (approval) → MEMORY.md
Weekly cycle	End of week	Reflection → coach review → SkillSchool analysis → trust recalculation
Digest	Daily (autopilot)	Batches notifications into a single summary

---

Stack

Layer	Technology
Backend	Python, FastAPI, WebSocket
Frontend	React, Vite, TypeScript, Zustand
Database	SQLite + sqlite-vec (local, optionally encrypted)
Embeddings	FastEmbed (local, never cloud)
Knowledge	Markdown vault (YAML frontmatter + wiki-links)
LLM	Anthropic / OpenAI or Ollama (local)
Desktop	Tauri (thin wrapper, system tray)

Project Structure

agntspace/
├── server/                    Python backend (FastAPI)
│   └── src/agntspace/
│       ├── api/               HTTP + WebSocket routes
│       ├── agent/             Main agent, orchestrator, subagent runtime
│       ├── llm/               Provider abstraction (cloud + local)
│       ├── vault/             Read/write/search markdown files
│       ├── journal/           Cognitive journal system
│       ├── tasks/             Task delegation + lifecycle
│       ├── trust/             Trust evolution engine
│       ├── skillschool/       Agent self-improvement
│       ├── coach/             Goal tracking + nudges
│       ├── skills/            Skill loader + matcher
│       ├── memory/            Conversation buffer + search
│       ├── domains/           9 Life OS modules
│       ├── channels/          Telegram, WhatsApp, Discord, Slack
│       ├── integrations/      Gmail, GCal, GitHub
│       ├── automation/        Heartbeat, scheduler
│       └── security/          Sanitizer, contract enforcement
├── ui/                        React frontend
├── desktop/                   Tauri desktop app
├── packages/plugin-sdk/       Python plugin SDK
├── defaults/                  Template vault + skills (copied on first run)
├── tests/                     Unit + integration + E2E
├── docs/                      Documentation
├── scripts/                   Dev, build, generators
└── deploy/                    Docker, systemd

User Data

~/.agntspace/                       ← app_dir (FIXED — config, db, credentials, logs)
├── config.toml
├── agntspace.db
├── credentials.json
└── logs/

<your folder>/                       ← vault root (any directory you pick)
├── agntspace-inbox/                 ← drop zone
├── agntspace-journal/               ← your journal + agent observations
│   ├── user/                        YOU write here (agent reads)
│   └── agnt/                        AGENT writes here (you read)
├── agntspace-knowledge/             ← knowledge bases (per topic)
│   └── {topic}/
│       ├── wiki/                    Compiled articles with [[backlinks]]
│       ├── library/                 Organized sources
│       └── inbox/                   Per-KB drop zone
├── agntspace-projects/              ← projects (tasks live inside each project)
│   └── {slug}/
│       ├── PROJECT.md
│       ├── tasks/
│       ├── notes/
│       └── output/
├── agntspace-goals/                 ← first-class goals (cross-project)
│
└── agntspace/                       ← agent internals (never touch manually)
    ├── system/
    │   ├── identity/                SOUL, USER, PROFILE, CONTRACT, TRUST, RELATIONSHIP
    │   ├── agents/                  Subagent definitions
    │   ├── skills/                  AI instruction sets
    │   ├── security/                Guardian, blacklist, audit
    │   └── logs/                    Server + simulation logs
    └── memory/
        ├── MEMORY.md                Long-term memory (assembled view)
        ├── summaries/               Layered temporal summaries
        ├── pending/                 Awaiting user approval
        └── graph/                   SPO knowledge graph

---

Status

Active development. Core systems implemented: chat, vault, LLM abstraction, journal, task delegation, trust evolution, contract enforcement, onboarding, dashboard, and control panel.

Extending AgntSpace

AgntSpace is proprietary software maintained by Archer Futura Technologies AB (Stockholm). The core product is not open source. The community hub is where you can extend it:

• Skills — teach the agent new behaviours with Markdown + YAML (no code)
• Plugins — register tools, hooks, and API routes in Python
• Vault templates — ship curated starter vaults for specific workflows

See docs.agntspace.com for authoring guides.

Support

• Bug reports / feature requests — community issue tracker
• Questions / show-and-tell — community Discussions
• Commercial support — agntspace.com

License

AgntSpace is proprietary software. By installing agntspace from PyPI you agree to the terms at agntspace.com/terms.

Free for personal use. Commercial licensing for organisations is available — contact [e-mail].