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.

page	what is on it
The Crucible	replication ledger — external claims rebuilt as the smallest runnable model, ruled REPRODUCED / FAILED / NOT_COMPUTABLE
Track record	forecasts scored against outcomes, including the ones we lost
Agent-Memory Integrity Leaderboard	does a correction actually stick? — the axis LoCoMo and LongMemEval do not test
inspeximus vs mem0 vs Graphiti	cross-system integrity comparison, run in each system's native configuration
Research writing	the essays, each one tied to a probe you can re-run

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

tool	one line	proof
inspeximus	agent memory + a self-maintaining second brain (value-ranked recall, consolidate, dead-link/orphan/stale repair)	python inspeximus/maintain.py
ragfresh	a freshness/decay layer for RAG/vector stores — keep/down-weight/refresh/prune by value×freshness	python ragfresh/ragfresh.py
nullcheck	is this number real, or noise? — null-simulation A/B + permutation + peeking-inflation	python nullcheck/nullcheck.py
selfref	is your AI training on itself? — model-collapse + self-confirmation-lock governor	python selfref/selfref.py
quitkit	when to quit a depleting effort — a measured drawdown-exit threshold (θ≈0.6)	python quitkit/quitkit.py
idcheck	is your causal/attribution number identified, or biased? — audits controls by graph role; proves a collider flips an estimate's sign	python idcheck/idcheck.py
goodhart	how gameable is your proxy/metric? — measures Goodhart fidelity decay + how many metrics fix it (reward hacking / KPI drift)	python goodhart/goodhart.py
herdcheck	will your multi-agent system herd? — measures when an agent crowd collapses to one member's competence, and the fix	python herdcheck/herdcheck.py

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

Method	Path	Description
GET	/api/v1/health	Server status + agent count
GET	/api/v1/agents/	List all active agents
GET	/api/v1/agents/{id}	Get agent by ID
POST	/api/v1/agents/{id}/pause	Pause agent
POST	/api/v1/agents/{id}/resume	Resume agent
POST	/api/v1/agents/{id}/reward	Reward agent (trust +)
POST	/api/v1/agents/{id}/punish	Punish agent (trust −)
POST	/api/v1/tasks/	Create task
GET	/api/v1/tasks/	List tasks
POST	/api/v1/tasks/{id}/assign/{agent}	Assign task
POST	/api/v1/tasks/{id}/complete	Complete task
POST	/api/v1/dungeon/spawn-agent	Spawn dungeon NPC
POST	/api/v1/dungeon/announce-task	Announce task (bidding)
POST	/api/v1/god/command	Execute God Console command (!help)
WS	/ws	WebSocket event stream

Environment Variables

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

Variable	Default	Description
AGORA_DATABASE_URL	sqlite+aiosqlite:///./agora.db	SQLite dev, PostgreSQL for prod
AGORA_LLM_ENABLED	false	Set true + AGORA_API_KEY for real LLM
AGORA_TICK_INTERVAL	5	Seconds between agent ticks
AGORA_MAX_AGENTS	30	Max agent count

Components

Directory	Tech	Purpose
server/	Python 3.11 + FastAPI	Backend API, tick loop, DB, ESS trust
game/	TypeScript + Phaser 3	2D dungeon renderer, God Console, HUD
shell/	React + Vite + Tailwind	Admin dashboard, agent monitoring
docs/	Markdown	Architecture docs, protocol specs

License

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