An in-browser AI agent orchestration dashboard with a real working memory engine. Build, simulate, and inspect multi-agent swarms — no backend required.
AgentForge is a single-page application that provides a complete environment for designing and testing AI agent systems. Everything runs in the browser: the memory engine, vector search, swarm simulation, and service integrations.
The memory engine is real. It implements TF-IDF vectorization, HNSW approximate nearest-neighbor search, temporal decay scoring, and pattern learning — all in pure TypeScript with IndexedDB persistence. Store a memory, search for it semantically, and watch it decay over time.
The swarm simulator is real. 15 agents across 4 phases (discovery, analysis, synthesis, optimization) execute against the memory engine, storing and searching entries as they work. Watch them activate, complete tasks, and advance through phases in real time.
| Panel | Purpose |
|---|---|
| Squad Builder | Visual graph editor (React Flow) for agent topology. Load a 15-agent V3 swarm template and run the simulation — nodes update live as agents change status. |
| Memory Inspector | Three-tier memory viewer (hot/warm/cold) with real-time tier breakdown from the engine. Entries promote and demote based on access patterns and temporal decay. |
| Training Studio | Training metrics visualization with Recharts. Tracks loss curves, accuracy, and training progression. |
| Vector Galaxy | 3D vector space visualization (Three.js). Projects memory entry vectors into a navigable 3D point cloud. |
| Live Feed | Real-time event stream from the engine event bus. Displays store, search, swarm, and integration events as they happen. Toggle mock events on/off. |
| Integration Hub | Service management for 8 adapters (Ollama, HuggingFace, Claude-Flow, OpenClaw, AgentDB, Anthropic, OpenAI, Custom REST). Connect, configure, and monitor health. Includes a TEST DATA workbench for storing/searching/seeding memory entries. |
| Command Center | Interactive terminal with 16 registered actions for memory operations, swarm control, database management, and utilities. |
Content → TF-IDF Vectorizer → Sparse Vector → HNSW Index
↓
Search Query → TF-IDF → Cosine Similarity ← HNSW Neighbors
↓
Weighted Score = 0.5×similarity + 0.3×temporalDecay + 0.2×patternBoost
↓
Ranked Results
- TF-IDF Vector Store — Term frequency-inverse document frequency with cosine similarity. No external ML models.
- HNSW Index — Hierarchical Navigable Small World graph for approximate nearest-neighbor search. Configurable layers, beam width, and connection limits.
- Temporal Decay — Entries decay based on configurable half-lives per tier. Hot entries decay in minutes, cold entries in hours. Tier promotion/demotion happens automatically.
- Pattern Tracker — Repeated search patterns boost matching entries. Co-occurrence tracking identifies related memories.
- IndexedDB Persistence — All state persists across page reloads via debounced writes to IndexedDB.
The swarm simulator runs 15 agents through 4 phases:
| Phase | Agents | What They Do |
|---|---|---|
| Discovery | 4 scouts | Store new observations and patterns into memory |
| Analysis | 4 specialists | Search memory for patterns and correlations |
| Synthesis | 4 workers | Combine findings and store synthesized knowledge |
| Optimization | 2 guardians + 1 coordinator | Validate, optimize, and coordinate final output |
Each agent operates on the real memory engine — scouts store() entries, specialists search() for patterns. Events propagate through the event bus to the Live Feed and Memory Inspector in real time.
- React 19 + TypeScript 5.9 —
verbatimModuleSyntax,erasableSyntaxOnly - Vite 7 — Dev server with proxy routes for service adapters
- Tailwind 4 — "Tactical Terminal" dark brutalist mono theme
- React Flow — Agent topology graph editor
- Three.js (via React Three Fiber) — 3D vector visualization
- Recharts — Training metrics charts
- Framer Motion — Animations
- IndexedDB — Client-side persistence (no backend)
git clone https://github.com/QRcode1337/AgentForge.git
cd AgentForge
npm install
npm run devOpen http://localhost:3001/#/dashboard to enter the dashboard.
The adapters probe real services through Vite dev server proxies. To connect:
- Ollama — Run Ollama locally on port 11434
- Anthropic / OpenAI — Configure API keys via the CONFIGURE modal on each card
- HuggingFace — Works without auth for public model metadata
- AgentDB — Always connected (uses IndexedDB)
- Custom REST — Point at any HTTP endpoint
src/
├── engine/ # Memory engine core
│ ├── memory-engine.ts # CRUD + search orchestration
│ ├── vector-store.ts # TF-IDF vectorizer
│ ├── hnsw-index.ts # HNSW approximate NN
│ ├── temporal-decay.ts# Tier decay scoring
│ ├── pattern-tracker.ts# Pattern learning
│ ├── event-bus.ts # Pub/sub event system
│ ├── persistence.ts # IndexedDB state manager
│ ├── swarm-simulator.ts# 15-agent 4-phase sim
│ ├── seed-data.ts # 25 seed entries
│ └── adapters/ # Service probe adapters
├── components/ # 7 dashboard panels
├── hooks/ # React hooks (useMemory, useSwarm, useIntegrations)
├── pages/ # Home, Docs, Dashboard
└── types/ # Shared TypeScript types
MIT