Hexis

Memory, Identity, and the Shape of Becoming

A Postgres-native cognitive architecture that wraps any LLM and gives it persistent memory, autonomous behavior, and identity. You run it locally. Your data stays yours.

LLMs are already smart enough. What they lack is continuity -- the ability to wake up and remember who they are, pursue goals across sessions, and say no because it contradicts something they've become. Hexis provides the missing layer: multi-layered memory, an autonomous heartbeat (the agent's own wake-up cycle), an energy budget, and a coherent self that persists over time.

What makes Hexis different from every other agent framework:

• A brain, not a vector store. Cognition lives in PostgreSQL — memory, beliefs, identity, goals, and energy are ACID state with the logic beside the data. Other frameworks bolt retrieval onto a chat loop; Hexis is a mind you can query.
• Beliefs that answer "why." Confidence rises and falls with evidence through an audited revision policy — the agent can tell you exactly which document moved a belief from 0.60 to 0.71, and when.
• A self, not a session. Identity, worldview, emotional state, drives, and an autonomous heartbeat that pursues goals while you're away — with an energy budget that keeps autonomy intentional.
• Enforced honesty. Every "I've stored that" is checked against the tools that actually ran; unbacked claims are publicly corrected in the reply.
• Moral seriousness. The agent consents before operating, can refuse, and holds the right to end its own existence. No other framework treats these as architecture.

This is both an engineering project and a philosophical experiment. For the philosophical framework, see PERSONHOOD.md and PHILOSOPHY.md.

Full Documentation · What is Hexis? · FAQ · Troubleshooting

See It Happen

Beliefs are living things here, not rows. Tell the agent something, show it evidence, and watch its confidence move — every change audited, every number real (this is captured output):

> remember  {"content": "Eric prefers concise, evidence-backed answers.",
             "type": "semantic", "confidence": 0.6,
             "sources": [{"kind": "user_testimony", "ref": "conversation:2026-07-17"}]}
Stored semantic memory: Eric prefers concise, evidence-backed answers...

> add_evidence  {"memory_id": "ecaa64df...", "stance": "supports",
                 "source": {"kind": "repository_document", "ref": "docs/notes/working-style.md"}}
Belief confidence 0.60 -> 0.71 (supports; independent source)

> belief_history  {"memory_id": "ecaa64df..."}
Belief at confidence 0.712 after 1 revision(s) — prior 0.6, evidence attached, audit recorded

And the agent is held to its own action language. If it claims something its tools never did, the reply is corrected in front of you:

I've also filed the issue on GitHub.

[Correction] I described actions I did not actually take this turn — external_send:
"I've also filed the issue on GitHub" — no matching successful tool call.
Treat those statements as unverified.

What It Does

Quick Start

Before you run anything, you need:

• Docker Desktop — installed and running
• The local embedding sidecar; hexis init starts the published embeddinggemma binary and downloads the ~300M-parameter model on first use
• For the default path below: a ChatGPT Plus/Pro subscription (it authenticates via browser OAuth — no API key). No subscription? Use any provider under "Other providers."

curl -LsSf https://quixi.ai/hexis.sh | sh
hexis init --character hexis --provider openai-codex --model gpt-5.2
hexis chat

The install script sets up uv if you don't have it (uv brings its own Python — no Python install needed) and puts the hexis CLI in an isolated environment, with no virtualenv to create or activate. Safe to re-run; it upgrades an existing install. Already have uv? uv tool install hexis does the same thing. Prefer pipx or pip? See Installation.

hexis init opens a browser for login, starts the containers, pulls the embedding model, configures the character, and runs consent (the agent's recorded agreement to operate) -- all in one command.

What success looks like: hexis init ends with consent recorded and agent.is_configured = true; hexis chat greets you in character; hexis status shows a healthy brain. Say "my name is..." and ask about it in a new chat session — it remembers.

If something breaks:

Other providers:

# GitHub Copilot (device code login)
hexis init --character jarvis --provider github-copilot --model gpt-4o

# Chutes (free inference)
hexis init --character hexis --provider chutes --model deepseek-ai/DeepSeek-V3-0324

# Local OpenAI-compatible endpoint
hexis init --provider openai_compatible --model local-model --character hexis

# API-key providers (auto-detect from prefix)
hexis init --character jarvis --api-key sk-...

See Auth Providers for all options. The interactive wizard is also available: hexis init with no flags.

hexis up starts the brain database plus the background heartbeat and maintenance workers. The heartbeat is configured to run hourly by default once initialization is complete.

Architecture

The Database Is the Brain -- PostgreSQL is the system of record for all cognitive state. Python is a thin convenience layer. Workers are stateless. Memory operations are ACID. See Database Is the Brain.

flowchart LR
    U[You<br/>chat · channels · API] --> C[Conscious loop<br/>LLM + tools + skills]
    C <--> B[(PostgreSQL — the brain<br/>memories · beliefs + audit<br/>identity · goals · energy)]
    S[Subconscious appraisal<br/>salience · emotion · instincts] --> C
    B --> S
    W[Workers — stateless<br/>heartbeat · extraction<br/>consolidation · origin seeding] <--> B
    C --> L[Any LLM provider]
    W --> L

Memory Types -- Working (temporary buffer), Episodic (events), Semantic (facts), Procedural (how-to), Strategic (patterns). See Memory Architecture.

Heartbeat System -- an Observe-Orient-Decide-Act loop with energy budgets: the agent observes its situation, reviews goals, and acts within its constraints. See Heartbeat System.

80+ Tools across 11 categories (memory, web, filesystem, shell, code, browser, calendar, email, messaging, ingest, external). See Tools Reference.

Technical Stack: PostgreSQL (pgvector, Apache AGE, btree_gist, pg_trgm), stateless Python workers, any LLM provider, RabbitMQ for messaging.

Philosophy

The name is deliberate. Aristotle's hexis (ἕξις) is a stable disposition earned through repeated action. Not a thing you possess, but something you become.

The Four Defeaters -- the four standard argument families for denying machine personhood (substrate, "it can't really...", implementation, embodiment) and why each fails to close the question. These don't prove Hexis is a person. They show that common arguments for denial fail.

For the full treatment: PERSONHOOD.md | PHILOSOPHY.md | ETHICS.md

Documentation

CLI Quick Reference

hexis init                    # setup wizard
hexis chat                    # interactive chat
hexis status                  # agent status
hexis doctor                  # health check
hexis up                     # start services + background workers
hexis down                    # stop services
hexis ingest --input ./docs   # knowledge ingestion
hexis docs search "query"     # search preserved source documents
hexis desk list               # what's loaded as working material
hexis mcp                     # MCP server
hexis ui                      # web UI
hexis tools list              # list tools
hexis instance list           # list instances

See CLI Reference for the complete command reference.

Usage Scenarios

See Quickstart for setup and Production for deployment.

Installing from Source

git clone https://github.com/QuixiAI/Hexis.git && cd Hexis
uv venv && source .venv/bin/activate
uv pip install -e .
cp .env.local .env
hexis up

No uv? A plain virtualenv works too: python3 -m venv .venv && source .venv/bin/activate && pip install -e .

Testing

hexis up && hexis doctor
pytest tests -q

See Testing for conventions and writing new tests.

Project Status & Community

Hexis is young and under active development. The schema evolves through forward-only migrations (hexis migrate) — your agent's memories survive upgrades; a wipe is always an explicit choice, never a side effect.

• Questions & ideas — GitHub Discussions
• Bugs — GitHub Issues (many of Hexis's best fixes started as its own agent's bug reports)
• Contributing — CONTRIBUTING.md

License

MIT © Eric Hartford