Agent-Memory: Persistent Memory for AI Coding Agents on iii

Overview

Agent-Memory is a persistent memory system for AI coding agents, built on the iii engine. It aims to stop agents from losing project context between sessions by capturing work and making relevant observations searchable later. Its integrations include Claude Code, Cursor, Gemini CLI, Codex CLI, and other agents that support MCP or HTTP, with one server sharing memory across clients.

The project is aimed at developers who use agents across repeated sessions or tools and want automatic capture rather than manually maintaining notes. It combines keyword, vector, and knowledge-graph retrieval; the README describes a local embedding option, but the underlying engine is iii.

Key Features

• Captures prompts, tool activity, and session events through agent hooks.
• Searches stored observations using BM25, vector search, and graph traversal, with results combined through rank fusion.
• Organizes memories into working, episodic, semantic, and procedural tiers, with decay and forgetting mechanisms.
• Provides MCP tools and a REST API for agent integrations, plus hooks and plugins for selected agents.
• Includes a local web viewer for sessions, observations, memories, and the knowledge graph.
• Offers team memory, provenance, and memory snapshots alongside multi-agent coordination features.

Use Cases

• A developer returning to a codebase after a break can retrieve prior decisions, fixes, and project context instead of restating them.
• A team using different coding agents can share project memory through the same server and MCP interface.
• Users of Claude Code or Codex CLI can automate session capture with the supplied plugins and lifecycle hooks.
• Developers who want to inspect what an agent has recorded can use the viewer to browse sessions and memories.

Getting Started

Requires Node.js 20 or later. Start the server with:

npx @agentmemory/agentmemory

Then connect an agent, for example:

agentmemory connect claude-code

See the repository README for agent-specific setup, configuration, and deployment details.

Alternatives

• a-mem-mcp: A-MEM MCP organizes coding-agent project knowledge as a self-evolving graph in ChromaDB, while Agent-Memory captures sessions through iii and shares retrieved context across clients.
• agent-memory: This TypeScript system turns session notes into tiered memory and generates AGENTS.md context, while Agent-Memory captures activity through iii and retrieves context across clients.
• plur: PLUR stores shared agent memory in readable local YAML files, while Agent-Memory uses iii to capture sessions and retrieve context across coding-agent clients.

Considerations

• The package requires Node.js 20 or later and runs on the iii engine. The README describes installing the engine separately or using Docker Compose as a fallback, so account for that runtime in deployment planning.
• The project’s measured status is Quiet: there have been no commits in the last 90 days, no releases in the last year, and no recent issue closures or merged pull requests. Most of the code comes from one person, despite a broader contributor list; this creates maintenance and continuity risk.
• The README notes that the project pins iii-engine to v0.11.2 because later engine versions require a migration that Agent-Memory has not completed. Check engine compatibility before upgrading either component.
• The README says LLM-backed compression and summarization are disabled by default. Some capabilities therefore require configuring a provider or explicitly opting into the documented fallback.
• The README reports benchmark results, but these are project-provided claims; validate them against your own agent workflows and data before relying on them.