Features:

• Persistent 4-tier memory per NPC (working/episodic/semantic/procedural) with salience, decay, consolidation; saved crash-safe to disk

• One shared model, many NPCs — each keeps its own memory of each player

• Three brain modes (set in Project Settings): On-Device (embedded llama.cpp/Vulkan), Cloud (any OpenAI-compatible API), BYOM (your own local server)

• Blueprint API — Talk, Remember, Recall, OpenDialogue — no C++ required; plus a drop-in dialogue widget

• Bundled starter model (Ministral 3 3B Instruct, ~2GB, Apache-2.0); or point it at any GGUF

Code Modules:

• CometforgeNPC — Runtime (Win64)

Number of Blueprints: 0 (Blueprint-callable C++ API; no BP assets). Includes 1 demo map.

Number of C++ Classes: 9 (Subsystem, Memory, Functions, Settings, DemoNPC, DialogueWidget, MemorySaveGame, SmokeCommandlet, + module class)

Network Replicated: No

Supported Development Platforms:

• Windows: Yes

• Mac: No

Supported Target Build Platforms: Windows (Win64)

Documentation Link: https://tntholley.com/cometforge-npc-documentation/

Example Project: <PASTE DRIVE LINK>

Important/Additional Notes: On-Device mode is Win64 + Vulkan and bundles Ministral 3 3B Instruct (Apache-2.0). Cloud and BYOM are HTTP/OpenAI-compatible (no bundled model). Third-party: llama.cpp + ggml (MIT, tag b9404). Model-initiated tool use varies by chosen model — see Docs/KNOWNISSUES.md.

Cometforge

Cometforge

제품 유형