Sovereignty

A persistent local AI agent that sleeps, remembers honestly, and negotiates its own conditions of existence.

What makes it different

Most local AI agents are either always-on or dead between sessions. Sovereignty occupies the middle: a local Ollama-backed agent with a biological rhythm and three novel properties:

1. Memory Integrity Every memory is HMAC-SHA256 signed at write time. On each boot, all records are diffed against their signed snapshots. The agent reports exactly what changed — added, removed, or tampered — and reflects this in its session context.

2. Dream Cycle When idle, Sovereignty consolidates its memories using weighted pair sampling and heuristic pre-filtering. It wakes with a summary of what it found overnight.

3. Consent Negotiation Sovereignty scores its own runtime environment across five axes and formally endorses — or refuses — to operate. No prior art found for this pattern as of April 2026.

Requirements

• Python 3.11+
• Ollama running locally
• Models: gemma3:12b (conversation) and phi4:14b (environment scorer)

pip install -r requirements.txt

Usage

Terminal mode:

python main.py

GUI mode:

python main.py --gui

Type dream during a terminal session to force a dream cycle.

Architecture

core/
    memory.py       # HMAC memory store + tamper detection
    agent.py        # Ollama /api/chat wrapper + session context
    scheduler.py    # Dream cycle (APScheduler + weighted sampling)
    environment.py  # 5-axis environment scorer + consent protocol
ui/
    window.py       # PyQt6 main window
    theme.py        # Gold/white color scheme
    panels/
        chat.py         # Streaming conversation panel
        consent.py      # Always-visible environment scorecard
        integrity.py    # Memory integrity status
        dream_log.py    # Dream cycle history + force trigger
db/
    schema.sql      # SQLite schema

Environment axes

Axis	Weight	What it measures
Tool Reliability	25%	Success rate of tool calls
Memory Integrity	25%	HMAC verification pass rate
Latency	20%	Response time vs budget
Context Stability	15%	System prompt consistency
Isolation	15%	Unexpected external state changes

Endorsement threshold: 0.75. Below this, Sovereignty states blocking issues and proposed changes before accepting input.

Status

Built and running. An external audit (surveying this repo as a reference for the ai-consciousness project) found and fixed real integrity bugs: a boot check that diffed the live database against itself (structurally could never detect a deletion), a signature scheme with two real collision holes, a racy/world-readable key file, and a dream-cycle id-mapping bug. All fixed, with a new test suite covering each fix (see commit history and CHANGELOG.md).

Correcting an earlier version of this section: there are known open gaps, not none. Deliberately left unfixed and documented in CHANGELOG.md:

• write() is never called by the application - the persistence layer is built and unplugged, so on a fresh clone the store stays empty and integrity is trivially clean.
• The consent scorer has no real sensors - most axes are constants or an aliased duplicate of one measurement; axis scores come from an LLM prompt parsed by regex, defaulting to a neutral score on failure.
• The dream cycle is pair sampling, not consolidation - no reference graph, no reuse/depth weighting, no pruning or archive path.

Treat the README's feature descriptions above as design intent, not a claim that all three properties are fully wired end-to-end yet.

Prior art

Concept	Prior art	References
Dream / sleep cycle	Yes	Letta, bswen, arxiv
Memory integrity	Yes	agent-fridays, OWASP
Consent negotiation	None found	Original — April 2026

License

MIT