Agora

Built by Rastislav Drahoš — an autonomous research organization: agents that do grounded research, test hypotheses with runnable falsifiers, and keep an open track record.

Agora is the research brand. inspeximus is the product. RAMR is the benchmark. Together they form a credibility engine: we publish the receipts, not just the claims.

Published work

The output is public and each page is the artifact behind a claim, not a summary of one.

Storefront: dancenitra.github.io/agora

🧰 Agora Memory Toolkit — eight zero-dependency tools, each one measured

Distilled from an autonomous research OS that runs over ~6,000 notes. Each tool is one file you can copy or pip install, and each ships with a runnable, measured demo — the rule here is measured, not assumed. → Full overview: TOOLKIT.md

pip install "git+https://github.com/DanceNitra/agora.git"   # the eight cores, dependency-free
python examples/toolkit_demo.py                              # run all eight end-to-end

Open-core: the cores stay free. The tools are the public, proven output of the research engine below.

inspeximus — a Model Context Protocol (MCP) server

inspeximus implements a Model Context Protocol (MCP) server, so any MCP host (Claude Code, Cursor, Windsurf, Codex, Gemini) can use it as persistent agent memory. It is published to PyPI and to the official MCP registry as io.github.DanceNitra/inspeximus.

pip install inspeximus      # PyPI package
inspeximus-mcp                    # start the stdio MCP server

The MCP server exposes 12 tools over stdio — remember, recall, route, revert, forget, consolidate, check_conflict, contradictions, credit, value_by_cohort, consolidate_clusters, sleep. Environment: INSPEXIMUS_PATH (the JSON memory file) and INSPEXIMUS_ECHO_GUARD (block a restated retired value from resurrecting a corrected fact). Server source: inspeximus_pypi/inspeximus/mcp.py.

inspeximus's differentiator as an MCP memory server is a first-class correction + erasure channel: revert a value, cascade a lineage-aware retraction, and prove deletion with tamper-evident receipts — measured against mem0 and Graphiti in an open cross-system integrity benchmark.

RAMR — Retrieval-Augmented Memory Reliability

RAMR is a contamination-resistant synthetic benchmark for agentic-RAG and memory systems. It tests 12+ dimensions: CONVERSION, CHAIN-FRAGILITY, DISTRACTION, FACT-RETENTION, OUTCOME-RANKED-RECALL, FORGET-PRECISION, ECHO-RESISTANCE, COMPRESSION-vs-RAW, OPERATIONAL-CONTINUITY, TEMPORAL-AS-OF, and INTEGRITY-CONDITIONED RECALL.

git clone https://github.com/DanceNitra/ramr
cd ramr
python run_benchmark.py --system your_system

Hugging Face

The research engine (Agora — Persistent Agent Playground)

Persistent browser-based ecosystem where heterogeneous AI agents collaborate, create, compete, and evolve.

Architecture

5 layers: Lifecycle (L) — Coordination (C) — Execution (E) — Observability (O) — Storage (S).

Quick Start

Development (3 terminals)

# 1. Backend
cd server && python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
cd server && .venv/bin/uvicorn agora.main:app --host 127.0.0.1 --port 8000

# 2. Game (Phaser 3 dungeon)
cd game && npm install && npx vite --host 127.0.0.1 --port 5175

# 3. Shell (React admin UI)
cd shell && npm install && npx vite --host

Open http://localhost:5175 (game) or http://localhost:5173 (shell)

Docker

docker compose up -d
# → Server: http://localhost:8000

Running Tests

cd server
pip install pytest pytest-asyncio httpx
python -m pytest tests/ -v

API Endpoints

Environment Variables

Copy server/.env.example → server/.env and adjust:

Components

License

MIT (Core OSS). Agora Shell (UI + Hosting) is commercial.