agent-memory: Give Coding Agents Persistent, Tiered Memory

Overview

agent-memory is a standalone memory system for AI coding agents. It collects session notes, uses an LLM to consolidate them into durable knowledge, and generates an AGENTS.md section so relevant context can be available at the start of a new thread.

It is intended for developers whose agents lose useful context between sessions. Its tiered disclosure model puts frequently relevant material in context, while keeping less relevant notes available in storage rather than requiring a keyword search for every fact.

Key Features

• Captures session notes in a journal queue for later processing.
• Consolidates journal entries into reusable topic-based memory with an LLM.
• Reorganizes entries through defragmentation, including merging duplicates and assigning hot, warm, or cold tiers.
• Generates AGENTS.md content with hot memory inlined and warm memory listed.
• Uses stable IDs and [[id__...]] links so entries can be cross-referenced after renaming.
• Provides a CLI for capture, listing, reading, consolidation, defragmentation, generation, and health checks.
• Supports adapter-based integration, with documented Amp and generic shell adapters.
• Stores memory as files, with a filesystem persistence implementation.

Use Cases

• An individual developer can carry project conventions and discoveries across coding-agent threads without manually searching old session notes.
• A team maintaining shared agent context can consolidate recurring decisions and technical knowledge into generated AGENTS.md content.
• Developers using different agent harnesses can use the standalone CLI or provide integrations through the adapter interfaces.
• A project that accumulates many memory notes can periodically reorganize them and disclose the most relevant subset to its agent.

Getting Started

Clone the repository and install dependencies with Bun:

git clone https://github.com/axiomhq/agent-memory
cd agent-memory
bun install

See the README for CLI commands, configuration, and integration options.

Alternatives

• ai-memory: ai-memory stores agent activity in a self-hosted, git-backed Markdown wiki, while agent-memory organizes session notes into tiers and generates AGENTS.md context.
• byterover-cli: ByteRover centers on locally curated, queryable, versioned project knowledge with optional cloud sync, rather than generating a tiered AGENTS.md section.
• cq: cq provides a workflow to query, contribute, and validate reusable coding lessons, while agent-memory turns session notes into tiered knowledge for AGENTS.md.
• TencentDB-Agent-Memory: TencentDB Agent Memory shares curated context and skills from conversations, documents, and code, while agent-memory focuses on coding-session notes and AGENTS.md.

Considerations

• Consolidation and defragmentation rely on an LLM adapter. Configure a command such as the documented amp agent run or integrate a compatible adapter for your environment.
• The documented standalone workflow uses Bun. The project also describes Nix flake and git submodule consumption.
• Memory is stored in files, so deployments need to account for where those files live and how they are shared or backed up.
• The repository is small, with 7 stars and no forks in the supplied project facts. Assess its fit and integration needs before depending on it for critical workflows.