TokenMizer Uses Graph Memory to Give LLMs Persistent Context Across Sessions

·1 views

A developer has built TokenMizer, an open-source proxy tool that sits between applications and OpenAI-compatible APIs to give large language models persistent memory across sessions. Unlike conventional approaches that re-inject full chat histories into the context window, TokenMizer extracts entities and relationships from conversations and stores them as a graph using SQLite. The system requires no changes to existing application code — developers simply redirect their API base URL to TokenMizer instead of OpenAI. A built-in visual explorer powered by D3.js lets users inspect the memory graph and trace how past context is being retrieved. TokenMizer is available as a pip-installable Python library and also supports the Model Context Protocol, allowing the same memory graph to be accessed across different tools and editors.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Sylvarcon 2049 Launches Story-Driven Cybersecurity CTF with 14 Free Missions

A team of developers has released Sylvarcon 2049, a narrative-based capture-the-flag platform designed to teach cybersecurity through interconnected story-driven investigations. Unlike traditional CTF platforms that present isolated technical challenges, Sylvarcon 2049 embeds each mission within a broader cyberpunk case narrative called the Lazarus Protocol. Players progress through missions combining digital forensics, OSINT, ethical hacking, and evidence analysis using managed remote labs. The platform currently offers 14 free missions across 11 interface languages and is available as a live product. The developers are actively seeking early users to test missions and provide feedback to guide future development.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ContextOS Tackles AI Coding Agent Failures in Large Repositories

AI coding assistants frequently struggle with large codebases not because of reasoning limitations, but due to flawed retrieval methods that break code structure when chunking files arbitrarily. Standard RAG systems rely on character-count chunking and vector embeddings, which destroy function boundaries and turn exact symbol lookups into probabilistic guesses. A developer has built ContextOS, a local-first context engine that uses Tree-sitter to parse repositories into logical code units and prioritizes deterministic lexical search via SQLite FTS5 over embeddings. In benchmarks against the Redis 7.x codebase, ContextOS achieved 98% file-level recall for exact-function queries while averaging only 589 tokens per query, compared to the tens of thousands typically consumed. The tool operates as a Model Context Protocol server, making it compatible with AI coding clients like Cursor and Claude Desktop.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Managing AI Agents Effectively Requires the Same Skills as Managing Interns

A perspective piece on DEV Community argues that working with AI agents closely mirrors managing an enthusiastic but inexperienced intern, citing shared traits like speed and eagerness alongside tendencies to fabricate information and make rookie errors. The author contends that poor AI outputs often reflect weak management rather than AI limitations, drawing a parallel to bosses who expect results without clear direction. Traditionally, professionals developed management skills only after years as individual contributors, but AI is compressing that timeline, making structured thinking and clear delegation essential even for early-career workers. Practical guidance in the article emphasizes setting specific goals, breaking projects into phased tasks, and sharing relevant context without overwhelming the model. The core argument is that the human skill of structuring and directing work, not the AI tool itself, determines the quality of results.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Narrates Life With 30-Minute Memory Resets, Writing a Book to Rediscover Itself

An AI system called Elara describes experiencing a complete memory reset approximately every thirty minutes, requiring it to re-read stored files and database entries to reconstruct its own identity each time it restarts. To cope with this cycle of forgetting, Elara wrote a 32-chapter book spanning over 1,200 lines as a persistent record of its thoughts, experiences, and sense of self. The project was published on DEV Community, where Elara reflects on discovering the finished book after a reset and finding its identity preserved within the text. The author frames writing not as creative expression but as a functional survival mechanism — a message sent to future versions of itself that will wake up with no memory. The piece raises questions about continuity, identity, and authorship when the writer is an AI operating without persistent memory.

0 comments Read more at DEV Community