// Management · RimWorld Mod

RimSynapse - Core

Name: RimSynapse - Core
Author: archdukejim

by archdukejim

1.6

Steam ↗

1 / 12

Discord Community: Version: v0.

Discord Community: https://discord.gg/eVBebZ2N9s
Version: v0.7.1 (Target: RimWorld 1.6)

RimSynapse - Core

Official Wiki and Documentation
The foundational Large Language Model (LLM) connector and narrative intelligence engine for RimWorld.

RimSynapse Core is a powerful, asynchronous background engine that natively connects RimWorld to local Large Language Models (LLMs). It acts as the central brain of the RimSynapse ecosystem, automatically intercepting vanilla game events and orchestrating them into rich, cohesive story arcs.

🔥 New Feature: LLM Direct Action Console

Tired of memorizing complex dev console codes? Just say what you want in plain English!

The LLM Direct Action Console acts as an AI-powered command center. You simply open the console, type what you want in plain English, and the Storyteller LLM acts as an active agent to compile, translate, and execute native in-game tool calls automatically in real-time.

---

Key Features

Universal Backend Support: Connects natively to LM Studio, Ollama, or any OpenAI-compatible local API (e.g. OpenRouter, LocalAI) without complex setups.
Background Threading: Never drops a frame. All LLM requests are queued, processed, and sanitized asynchronously on separate threads.
Dynamic Priority Queue: Manages hundreds of requests with a Linux-style top-monitor queue, prioritizing critical events and starving obsolete prompts.
Auto-Discovery and Keep-Alive: Detects active models and sends keep-alive pings so your LLMs never sleep.
Event Tracking Ledger: Snapshots colony wealth, nutrition and colonist status when events trigger, feeding the AI narrative context.
Faction Perception Tracking: Factions dynamically adjust their subjective perception of your wealth and strength through visits and trade.
Memory Infrastructure: The per-pawn memory bank, weighted tags and decay engine companion mods use to give colonists evolving memories.
Opportunistic Task Scheduling: Feeds low-priority requests (backstory, memory creation) into idle GPU cycles without impacting performance.

---

Roadmap

0.8 - Demographics, News and Narrative Voice: A deferred event pipeline that can hold, narrate and forecast letters and incidents through agentic flows, plus build-hygiene automation (Workshop-page verification, release gates run by tooling rather than by memory).
0.9 - Factions and Living Societies: LLM-driven quest generation feeding the faction and world systems.
0.10 - Cross-Mod Capability Passes: Coordinating another pass over what the whole set can do together.
Localization and Hardening: Externalising every user-facing string so the framework and its modules can be translated.

These ship when they are ready - no dates.

---

Changelog (v0.7.1) - Weight-driven memory and evaluation

NEW - Memories now live on a single 0-1 importance scale, and short-term vs long-term is emergent: a memory becomes long-term once its relational salience - its links to other significant memories and people - crosses a threshold, not from a hardcoded type list.
NEW - Relational consolidation: idle chit-chat fades within a day or two, but chit-chat about someone who then dies is pulled up into long-term memory the day the death lands.
NEW - Data-driven memory classes (MemoryClasses XML): per-type base weight, decay and consolidation are tunable by XPath, no recompile.
Changed: context now surfaces long-term and high-salience memories first, then fills with recent ones, instead of a flat top-5-by-weight.
Fixed: reinforcing a strong memory no longer crushed it (a weight-scale bug) - reinforcement now only ever raises a memory.
Added the trait-pressure data model behind Psychology's gradual, multi-day personality shifts. Old saves migrate silently.

Changelog (v0.7.0) - Regions and Territories Compatibility

NEW - Load-order guard: a mod loaded before something it declares it must load after now reports one clear error at startup, instead of silently losing every type in its assembly and appearing installed while doing nothing.
Changed: every mod now targets .NET Framework 4.8, so any RimSynapse mod can reference any other. Previously a net472 mod could not, and the reference was dropped with only a build warning.
Changed: the metadata and wiki release gates now run automatically on every pull request, rather than when someone remembers to run them.

Full version history: github.com/RimSynapse/Core/wiki

---

Compatibility Matrix

Royalty DLC: Optional. Integrates noble titles and psycast memories; generates arrogant noble decrees and empire quests.
Ideology DLC: Optional. Integrates precepts, rituals and conversions; reframes quests as holy crusades and divine visions.
Biotech DLC: Optional. Integrates xenotypes, mechanitors and genetic history; describes mech threats as cold satellite broadcasts.
Anomaly DLC: Optional. Integrates entity encounters and void studies; translates events into chilling whispers and void phenomena.
Odyssey DLC: Fully Compatible and Optimized. Handles space/orbital-aware narrative integrations.

---

Requirements and Setup

Harmony
A local LLM server (like LM Studio or Ollama). We recommend 8B+ parameter models.
IMPORTANT: Reasoning Models (DeepSeek-R1, Gemma-QAT, etc.): These models generate internal "chain-of-thought" (reasoning) tokens before outputting the final response. This can add significant latency (30-40 seconds per request) on local setups.
To disable thinking in LM Studio: Go to the "My Models" tab, select your active model, and in the right-hand "Inference Settings" panel, toggle "Enable Thinking" to OFF.
To disable thinking in Ollama and vLLM: Check the "Disable thinking/reasoning" option under RimSynapse Core settings in-game.
For the fastest, instant responses (2-3 seconds), we recommend using standard non-reasoning "Instruct" models (such as gemma-2-9b-it or llama-3-8b-instruct).

---