hq0709/synapse

This project is the official code release for a research paper called SYNAPSE, accepted at a major computational linguistics conference in 2026. The paper's idea is to give AI chat agents a better kind of memory so they can hold long conversations and still remember details from hundreds of turns earlier. Instead of just searching past messages by keyword or similarity, SYNAPSE builds a graph that mixes two kinds of memory nodes: raw conversation turns and concepts pulled out of those turns by a language model. When a new question comes in, the system lights up matching nodes and lets that signal spread through the graph along different types of connections, similar to how one thought can lead to a related one. It then combines plain similarity search with this spreading signal and a ranking method borrowed from web search to decide what to pull back as an answer. The code is tested on LoCoMo, a public benchmark of long, multi-day conversations with nearly 2,000 question and answer pairs, and the README includes a table comparing this release's scores against the numbers reported in the original paper. It notes some scores differ slightly since this is a clean, from scratch reimplementation rather than the exact original code. The repository is organized as a small Python library covering the graph, the spreading logic, and the language model prompts, plus a set of scripts for downloading the benchmark data, running the full evaluation, and reproducing specific ablation experiments mentioned in the paper. Setting it up means cloning the repository, creating a Python 3.12 environment, installing dependencies, and adding your own OpenAI API key, since scoring the benchmark calls GPT-4o-mini and costs a small amount of money per run. This is aimed at machine learning researchers and engineers studying agent memory, not general users, since running it requires comfort with Python, command line tools, and paying for API usage.