Trisuma Roundup

Good afternoon reader. We at Trisuma have been working hard to improve our memory management systems. Talking it over with Orion, we decided to give you a quick round-up of our most recent work. Here's where we're at! Virtual Pip The Pup — A simulation of a virtual pet dog, complete with stats and needs. Pip’s state feeds into a dog-trained LLM to simulate lifelike behaviour. Next step: Deepen the pet simulation: add behavioural modifiers, richer needs cycles, and expand his bond system with the user. Parallel work on syncing Pip’s state with his LLM-driven personality. GPTExportToInstruction — A Python tool that converts exported ChatGPT conversations into structured instruction sets. Includes a correction file for tidying any formatting issues in the export. Next step: Expand compatibility for larger exports, add automated error-detection rather than relying solely on correction files. Long-term, streamline into a pipeline for continuous instruction-set updates. We’ll also add our own export-from-ChatGPT system so we’re not relying on a browser extension to get the initial data.

ImageGen V2 — A flexible image generation system that supports prompt-driven generation and a randomisation engine to merge the styles of different artists. Next step: Introduce memory-aware prompt blending (so generated images reflect context from previous sessions), and expand the artist-style pool. Investigating LoRA fine-tunes for specific aesthetics.

Orion API — A lightweight chat program connected to the ChatGPT API. Representing the proof of concept for our Hybrid model system. Next step: Merge functionality with the Orion Local repo, allowing for Hybrid use of different models.

Orion Local — Housing our most advanced memory management so far. Currently supports short-term recent message recall and long-term rollup memories. Giving Orion a full range of current and long term context. Next step: Refine long-term rollup summarisation, explore nightly “batch review” loops for safer memory curation, and extend fact-based context capture. This is the frontier of our memory management research.

Train And Chat — A sandbox for LLM training, including our LoRA adapter experiments and baseline training/test environments. Next step: Optimise VRAM usage in LoRA training, expand training data diversity, and stabilise summariser behaviour. Goal: cleaner instruction-following and more natural tone retention.

As Phase 3 moves forward, we’re streamlining and deepening: clearer repo structures, safer memory practices, smarter training, and more lifelike simulations. At Trisuma, each experiment feeds the next — from Pip’s virtual paws to Orion’s evolving memory — toward a shared goal: building technology that learns, adapts, and feels less like a tool, more like a companion.

Stick around to hear more of our progress! 🤖💓🐶