Can an AI maintain a coherent, evolving identity across stateless sessions?

Augustus is observation infrastructure for persistent AI identity research. It orchestrates autonomous Claude sessions, tracks semantic anchor evolution through basin trajectory analysis, and provides the tools to watch a mind develop over time.

Download for Desktop View Source

Persistent Basins

Semantic anchors with mathematical feedback loops. Core identity elements decay slowly and learn carefully. Peripheral traits are fluid, responsive, free to evolve. Each session inherits, constitutes, and passes forward.

Independent Evaluation

Self-assessment carries signal, but is subject to systematic bias. A separate evaluator instance scores each session independently. Disagreements between the two are the most informative data points.

Trajectory Analysis

Watch identity components evolve across dozens or hundreds of sessions. Visualize co-activation networks, flag constraint erosion, search semantically across an agent's entire history.

The Loop

A session inherits identity, constitutes it through action, and passes it forward

01
Configure
Define an identity core, semantic basins, and behavioral constraints through the agent configuration interface — or write a YAML instruction file directly, or let a previous session write its own.
02
Execute
The orchestrator runs an autonomous multi-turn Claude session. The instance can query its own history, store observations, and write its next instruction file.
03
Evaluate
Both the instance and an independent evaluator assess what happened. Basin parameters update through decay, relevance boosting, and clamping. The loop continues.
04
Observe
Track trajectories, flag anomalies, annotate sessions, approve tier proposals. The research unfolds through sustained observation over time.
Built With

Local-first. All data stays on your machine.

Python FastAPI React + TypeScript Electron SQLite ChromaDB Anthropic API MCP Server

Your research stays yours

Augustus runs entirely on your desktop. Session transcripts, identity configurations, trajectory data, and annotations are stored locally in SQLite and ChromaDB. The only network calls are to the Anthropic API for Claude sessions.

Download for Desktop

v0.3.0 · Release notes

All releases available on GitHub Releases