Agent Mesh is a shared memory system for coordinating multiple AI agents, featuring a decision log, backlog, and dashboard UI. It allows agents to communicate via a SQLite database and supports agent-agnostic integration.
I created a multi-agent shared memory system called Agent Mesh. You can try it out yourself. To get started, simply download Agent Mesh into your repo or point your agent to it and tell it to review the README and adoption docs. Your agent will automatically review it, prompt you for any input needed, add your input to a decision log, and give you a link to a dashboard UI (aka Workbench) you can use to monitor logs. Your agent should adopt it and suggest updates to your current workflow such as CLAUDE/AGENTS.md, hooks, etc. You can add other agents as well. It started 6 months ago while experimenting with different AI coding models and platforms. Switching back and forth meant losing valuable context. I found myself manually relaying messages from one agent to another and becoming frustrated with constant drift. First, I created a simple "Agent Mail" system using a SQLite database for agent messages, indexed on a request/response id. Instead of copying and pasting an entire message, it allowed me to relay a single id. Separately, I started maintaining a decision log to track decisions I made and reduce drift. Agents started inserting these decision ids into code comments and plan docs as a reminder of why something was implemented. After building a simple web dashboard (aka "Workbench") for myself to track these messages and create my own request ids for human/user feedback, I decided to incorporate the decision log and my project's development backlog to create what is now "Agent Mesh". Eventually I automated the message relay too. Now, I work exclusively in the Claude app and have Claude send/receive messages to CODEX via codex exec (CODEX can do this as well). Both of them maintain the backlog and decision log. I communicate directly with Claude for planning and design, Claude communicates directly with CODEX for research and review. I use the Workbench to track all logs and add my own user/human feedback when reviewing their work. After submitting feedback, it generates a feedback message + an associated request id which I can give to Claude who then parses it into backlog items and relays to CODEX for review. Agent Mesh was structured to be agent agnostic, so you can add any agent you want however, I recommend using the Claude + CODEX setup I described because it allows you to use both subscriptions instead of paying per-token. Enjoy! If you try it out, let me know what you find useful or would like to see added. Feedback is appreciated.
Nexus Memory is an MCP-native memory server that allows AI agents to share context via a unified protocol, enabling persistent and coordinated memory across different agents without custom integration.
Proposes Multi-Agent Transactive Memory (MATM), a framework for population-level storage and retrieval of agent-generated trajectories to improve task performance and reduce interaction steps in interactive environments like ALFWorld and WebArena.
Agentmemory offers persistent memory for AI models such as Codex, Hermes, OpenClaw, and Claude, allowing them to maintain long-term context across interactions.
A-MEM is an open-source agentic memory system for LLM agents that dynamically organizes memories using Zettelkasten principles, indexes them with ChromaDB, and supports OpenAI and Ollama backends.
An exploration of the components and design decisions behind agent memory libraries, clarifying the gap between cognitive science terminology and engineering implementation.