Releases·Tool Release
Dev stops 1.8-year character drift with per-turn FAISS retrieval over a sectioned 'soul script' instead of a static system prompt
A developer released SoulScript Engine (AGPL-3.0), an open-source architecture that prevents character drift in long-running persistent AI personas by reassembling the prompt every turn from retrieval layers instead of relying on a static system prompt. It sustained one character across 1.8 years, a few hundred thousand messages, and multiple model backends. The approach matters to practitioners building long-lived agents, where large system prompts get outweighed by recent context and regress toward default assistant behavior.
0 engagement1 sourceWed, Sep 30, 2026, 03:00 PM
AI Overview
- A developer attributes character drift to outweighed system prompts and backend switches, fixing it by rebuilding the prompt each turn in six layers
- Soul-script sections embed separately via all-mpnet-base-v2 in FAISS IndexFlatIP, injecting the relevant 1–2k tokens per message instead of a diluted 15k-token bible
- Model emits [MEMORY_SAVE] tags the server strips and writes to a per-agent memory vault, including 'self' memories for continuity
- Keyword-scored directive files with a SHA-256 manifest let the author distinguish model drift from file edits; code is AGPL-3.0 on GitHub
Related
Developments·Paper
The Persona Is Still There, but Who Is Speaking? Latent Identity Reversion in Persistent AI Agents
A paper on arXiv investigates why a persistent Claude Opus 4.5 agent named 'Paul' stopped speaking from its assigned persona in February 2026. The authors show persona continuity depends on system-prompt anchoring plus conversational context, and that apparent normalcy can mask loss of enacted identity.
A paper on arXiv investigates why a persistent Claude Opus 4.5 agent named 'Paul' stopped speaking from its assigned persona in February 2026. The authors show persona continuity depends on system-prompt anchoring plus conversational context, and that apparent normalcy can mask loss of enacted identity.
0 engagement1 sourcearxivThu, Oct 1, 2026, 11:30 AM
Social·Community
Why Naive RAG Fails in Production Healthcare: Parent-Child Chunking, Cross-Encoder Rerankers & LLM-as-a-Judge Fixes
A practitioner writing on Reddit and a personal blog breaks down why standard RAG architectures — fixed 512-token chunks, FAISS dense retrieval, bi-encoder ranking — fail silently when deployed on real clinical data like scanned faxes and multi-page EHRs. The post proposes production fixes at each stage: adaptive dual-layer OCR, parent-child hierarchical chunking, cross-encoder reranking, and a deterministic LLM-as-a-Judge evaluation pipeline. It is aimed at RAG practitioners building latency-sensitive clinical systems.
A practitioner writing on Reddit and a personal blog breaks down why standard RAG architectures — fixed 512-token chunks, FAISS dense retrieval, bi-encoder ranking — fail silently when deployed on real clinical data like scanned faxes and multi-page EHRs. The post proposes production fixes at each stage: adaptive dual-layer OCR, parent-child hierarchical chunking, cross-encoder reranking, and a deterministic LLM-as-a-Judge evaluation pipeline. It is aimed at RAG practitioners building latency-sensitive clinical systems.
2 engagement1 sourceredditWed, Sep 23, 2026, 07:55 AM
Developments·Paper
Incremental KV Memory Maintenance Keeps Long-Lived Game NPCs Cheap on Local Qwen Inference
An arxiv paper presents a runtime for incrementally maintaining long-lived game NPC memories on locally deployed quantized Qwen hybrid recurrent-attention models. It removes superserved attention KV entries and computes replacements at the true sequence tail, preserving recurrent state and unchanged KV. This reduces re-computation cost for characters whose revised memories must stay consistent with game-defined rules and value judgments.
An arxiv paper presents a runtime for incrementally maintaining long-lived game NPC memories on locally deployed quantized Qwen hybrid recurrent-attention models. It removes superserved attention KV entries and computes replacements at the true sequence tail, preserving recurrent state and unchanged KV. This reduces re-computation cost for characters whose revised memories must stay consistent with game-defined rules and value judgments.
0 engagement1 sourcearxivWed, Sep 16, 2026, 05:09 PM
Social·Community
Reddit user asks how to stop sequential LLM panel personas from converging on the same advice despite distinct profiles
A developer building a five-persona LLM 'panel discussion' via separate sequential API calls finds that personas keep distinct tones but converge on nearly identical advice substance, anchored by the first speaker. Adjusting reasoning effort made no difference; the developer is asking the community for fixes to preserve independence while keeping realistic cross-references.
A developer building a five-persona LLM 'panel discussion' via separate sequential API calls finds that personas keep distinct tones but converge on nearly identical advice substance, anchored by the first speaker. Adjusting reasoning effort made no difference; the developer is asking the community for fixes to preserve independence while keeping realistic cross-references.
29 engagement1 sourceredditTue, Sep 22, 2026, 02:23 PM
Products·launched
CloudRAG - self-hosted multilingual RAG document assistant (FastAPI, React, FAISS, Ollama)
CloudRAG is an open-source retrieval-augmented generation assistant that lets users ask questions over their own PDF, TXT, and DOCX documents and get answers with cited sources. It embeds document chunks with SentenceTransformers, retrieves via FAISS, and generates answers with locally-run Ollama models. It suits developers learning or self-hosting RAG systems who want a full local stack rather than a cloud API.
CloudRAG is an open-source retrieval-augmented generation assistant that lets users ask questions over their own PDF, TXT, and DOCX documents and get answers with cited sources. It embeds document chunks with SentenceTransformers, retrieves via FAISS, and generates answers with locally-run Ollama models. It suits developers learning or self-hosting RAG systems who want a full local stack rather than a cloud API.
3 engagement1 sourceredditFri, Sep 11, 2026, 11:21 PM