agentmemory: Persistent Memory for AI Coding Agents

Overview

agentmemory is a persistent memory service for AI coding agents. It helps agents retain useful project context between sessions, reducing repeated explanations and rediscovery of prior decisions. Memory can be shared across compatible agents through a common server rather than kept in separate agent-specific files.

The project is aimed at developers who want automatic session capture and searchable recall, with integrations for coding agents that support hooks, MCP, or HTTP. It is most useful when work spans many sessions or agents; it may be unnecessary for projects where a small static instruction file is sufficient.

Key Features

• Captures prompts and tool activity through agent lifecycle hooks.
• Searches memory using BM25, with optional vector embeddings and graph-based matching.
• Consolidates observations into working, episodic, semantic, and procedural memory tiers.
• Provides MCP tools and a REST API for agent integrations and other clients.
• Offers a local real-time viewer and session replay.
• Includes integrations and setup paths for a range of coding agents.
• Supports keyless keyword search, with local embeddings available as an option.

Use Cases

• Developers returning to a codebase after a break can retrieve prior implementation choices and debugging context.
• Teams using several coding agents can share project memory through one MCP or REST service.
• Users of hook-capable agents can capture session activity automatically instead of maintaining notes by hand.
• Developers evaluating agent recall can use the included benchmark and evaluation materials to inspect reported results.

Getting Started

Requires Node.js 20 or newer. Start the interactive setup with:

npx -y @agentmemory/agentmemory@latest

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

Considerations

• The local runtime depends on the iii-engine, pinned by the project to version 0.22.1. On native Windows, the engine must be installed manually, or users can use WSL2 or Docker Desktop.
• Keyless setup supports BM25 keyword search but disables vector embeddings. Local semantic embeddings require an initial model download; other embedding and LLM features depend on configuration.
• Some README benchmark results are project-reported and use different datasets or baselines from competitor figures. Treat cross-project comparisons as indicative rather than direct head-to-head measurements.
• The project has substantial adoption, but the supplied metadata also lists 599 open issues. Review current releases and issue activity before relying on it in a critical workflow.