working-memory

We introduce temporal working memory (TWM), which aims to enhance the temporal modeling capabilities of Multimodal foundation models (MFMs). This plug-and-play module can be easily integrated into existing MFMs. With our TWM, nine state-of-the-art models exhibit significant performance improvements across QA, captioning, and retrieval tasks.

Personalized Training for the Sequence Learning task with the NAO robot and the MUSE EEG sensor

Spiking neuronal network simulations (Python, NEST Simulator) for continuous attractor working memory networks with short-term plasticity.

MCP Context Server — a FastMCP-based server providing persistent multimodal context storage for LLM agents.

A setup to evaluate working memory using EEG signals

MemGPT-inspired tiered memory system with working memory, archival storage, and automated context paging

MemGPT-inspired tiered memory system with working memory, archival storage, and automated context paging

Visual Working Memory Game in Python which gives you the gameplay data

SQLite of agent memory — single-file, embeddable, zero-config. Episodic + semantic memory, hybrid BM25+vector recall, bitemporal facts, spreading-activation, working memory, async API, LLM compression, Ebbinghaus decay, GDPR forget, multi-agent, MCP-native.

A unified working & short-term memory task for artificial neural networks.

This repository contains all codes necessary to reproduce figures and results reported in Stein, Barbosa et al. (Nature Communications, 2020) from the raw data acquired in human behavioral experiments (data included in the repository), and from the relevant model simulations.

A spiking network model of dopamine-gated working memory persistence in a ring attractor. Below a critical D1 gain level, the memory bump collapses intrinsically - not because of distractor interference.

Semantic codebase search + persistent working memory for AI code editors. Local, zero-config, MCP. No API key.

These are short example PsychoPy Scripts

Generative working memory in Transformer decoder

Digital implementation of the Corsi Block-Tapping Test (Forward & Backward) with trial-level logging.

Cognitive inspired Nengo simulation

Biologically-inspired Neural Network for lifelong learning — reference implementation of the IEEE CogMI 2025 paper on overcoming catastrophic forgetting with short-term, working, and long-term memory

A persistent, human-inspired memory system for LLMs built with Python, FastAPI & React. Implements all 4 cognitive memory types: working, episodic, semantic, and procedural. Features Ebbinghaus forgetting curve decay, auto memory consolidation using Gemini 2.5 Flash, and a live memory inspector UI.

Visual chain-of-thought reasoning via a step-by-step visual working memory — decompose, operate, then answer