Skip to content

🧠 MemoryGr.id: Individual & Collective Memory System for AI societies

MemoryGrid

MemoryGrid is a distributed memory system where agents manage their own layered memory systems but also contribute to and draw from a collective, evolving memory base, enabling both individual reasoning and coordinated collective intelligence.

Agent Level

At an agent level, MemoryGrid is a multi-layered collection of modular & interoperable memory systems, where each memory system is for specific types of memory (facts, experiences, skills, reflections, and plans) and all layers integrate into a cognitive memory whole.

Agency Level

At an agency level, be it at org or society, MemoryGrid is a vast memory library built & contributed to by many agents.

  • Each agent contributes its own select memories of knowledge, experiences, facts, and plans.
  • Every agent can then read, borrow, and update this shared collection, enabling both personal learning and collective reasoning.

Foundational Role

MemoryGrid is a foundational subsystem for multi-agent cognition.

  • Each memory type - semantic, episodic, procedural, reflective, and strategic - functions as a self-contained unit.
  • Yet, all connect through standardized interfaces to form a higher-level integrated cognitive memory system.

🎯 Supported Core Memory System Types

Memory Type Brief
Semantic Memory The structured knowledge store of facts, concepts, relations, and norms that agents use to reason and coordinate.
Episodic Memory The record of contextualized experiences - the what, where, when of interactions that anchor learning and trust.
Procedural Memory The store of skills, routines, and action patterns that enable agents to act fluently and reliably.
Working Memory The short-term cognitive workspace where information is actively combined, tested, and used for immediate reasoning.
Reflections Memory The meta-cognitive layer where agents evaluate past actions, extract insights, and refine future strategies.
World Model Memory The global cognitive map of reality, encoding causal structures and dynamics for prediction and planning.
Communication Memory The record of dialogues, interactions, and commitments that sustain continuity, trust, and cooperation.
Reward Memory The motivational compass that tags outcomes with value, shaping preferences, strategies, and adaptation.
Context Cache The fast, ultra-short-term buffer that keeps agents coherent and synchronized in real-time interactions.

⏳ Temporal & 🗂️ Representational forms of Memory

Memory in agents spans two axes: the temporal dimension (how long information lasts and is used) and the representational form (how information is structured and accessed). Together, these define the architecture that shapes intelligence’s depth, flexibility, and adaptability.

Dimension Type Brief
Temporal Short-Term Memory A volatile buffer for immediate context, signals, and transient states, enabling real-time reasoning and coordination.
Long-Term Memory A durable repository of knowledge, skills, and histories that forms the basis for cumulative learning, identity, and alignment.
Representational Vector Memory Encodes knowledge as embeddings for fast, similarity-based retrieval and statistical alignment across agents.
Tree Memory Organizes memory hierarchically, supporting structured reasoning, planning, and task decomposition.
Graph Memory Stores knowledge as networks of relations, enabling rich semantic reasoning, causal modeling, and shared ontologies.

🧩 Strategic Memory

Strategic memory encodes the directional logic of intelligence — storing and updating plans, goals, states, values, and norms. Unlike content-focused (episodic, semantic) or process-focused (working) memory, it organizes cognition around purpose and alignment, enabling agents to act with direction, consistency, and coherence over time.

Memory Type Brief
Plans Memory Stores multi-step strategies and sequencing logic, enabling agents to turn intentions into coordinated execution.
Goals & Task Memory Preserves long-term goals and short-term tasks, anchoring agent behavior in purpose-driven direction and prioritization.
State Memory Records internal and external conditions, providing situational context for adaptive reasoning and coordination.
Normative Memory Encodes rules, norms, and commitments, ensuring alignment, compliance, and fair coordination across agents.
Value Memory Holds ethical principles, cultural values, and preference systems, guiding decisions beyond immediate utility or rewards.

🛠️ Shared Memory Infrastructure

The collective substrate that enables agents to pool knowledge, goals, and signals for distributed sense-making and coordination.

Memory Type One-Line Definition
Shared Knowledge Base A common repository of facts and ontologies that prevents semantic drift and grounds agents in a consistent reference frame.
Consensus Memories The authoritative record of collective agreements and validated decisions, ensuring durable and enforceable commitments.
Global Workspace A system-wide broadcast channel for salient signals, aligning distributed attention and enabling rapid adaptation.
Blackboard System A modular collaboration surface for asynchronous contributions, enabling iterative integration of diverse expertise.

🏗️ Core Building Blocks of AgentGrid

The MemoryGrid is built upon the following key projects, each contributing a unique piece of the cognitive memory whole:

Project Intuitive Brief
🎮 OpenArcade Framework to shape agent populations; enables strategies for interaction, collaboration, cooperation, negotiation, and social decision-making.
🌐 OpenMe.sh Open, protocol-native communication mesh; enables signaling, message exchange, and shared context across groups, orgs, and geographies.
🔗 Pervasive.Link Meta-protocol that binds heterogeneous systems; encodes, translates protocols, context, languages, and strategies into interoperable structures.

MemoryGrid is fundamentally a grid of general-purpose, distributed & decentralized memory systems that are capable of indexing multi-modal data such as documents, video streams, high-resolution images, sensor data, AI inputs/outputs and any general data. It supports unified APIs over in-memory, persistent, and streaming backends, making it ideal for modern data-driven applications, AI pipelines, and real-time analytics.

MemoryGrid is Kubernetes-native, supports pluggable backends and custom serialization, and is designed for easy integration with ML/AI systems.

📚 Contents

🏗 Architecture Diagrams

🚀 Features

  • 🔁 Unified Read/Write API across Redis, TiDB, and stream queues
  • 📦 Object abstraction with typed backends: in-memory, storage, stream
  • 🧠 Custom Serialization support (e.g., pickle, JSON, custom codecs)
  • 💾 Backup & Restore support for external object stores (S3, GCS, MinIO, Ceph)
  • 🌐 Multi-cluster routing for object and stream metadata
  • ⚙️ Kubernetes-native deployments with dynamic service provisioning
  • 💡 AI Integration Ready: store Python objects (e.g., ML model snapshots, configs)
  • 📽️ Built-in GStreamer pipelines with GPU acceleration for video ingestion

🧠 Core Concepts

📂 MemoryGrid Instances

Type Backend Use Case
In-memory Redis Fast access, low latency data
Persistent TiDB Durable, queryable blob storage
Streaming Redis Producer-consumer style pipelines

🔧 MemoryGrid Services

  • Config Service: Creates and manages DB instances across clusters
  • Routing Service: Tracks object/stream metadata and their locations
  • Writer Client SDK: Python interface for storing, retrieving, and streaming objects
  • Object Service: gRPC interface for Set/Get/Stream operations
  • Video Ingestion: GStreamer-based GPU pipelines for real-time decoding

🎥 Video Ingestion with GStreamer

MemoryGrid includes GPU-accelerated pipelines for ingesting video streams in both live and stored modes. Pipelines are deployed as containers and exposed via REST endpoints.

Mode Usage
Live Real-time RTSP camera ingestion
Stored Archived video file playback

📢 Communications

  1. 📧 Email: community@opencyberspace.org
  2. 💬 Discord: OpenCyberspace
  3. 🐦 X (Twitter): @opencyberspace

🤝 Join Us!

This project is community-driven. Theory, Protocol, implementations - All contributions are welcome.

Get Involved