claude-mem gives coding agents a memory that lasts across sessions

What this is about

Anyone who works with coding agents such as Claude Code, Codex or OpenCode knows the problem: every new session starts almost from zero. The agent no longer knows which bug fixes worked yesterday, which database decision was made, or which approach already failed. You explain it again, paste notes into the prompt or maintain files such as CLAUDE.md by hand.

This is where claude-mem comes in, an open-source project by Alex Newman (GitHub account thedotmack). The repository is licensed under Apache 2.0 and, according to the GitHub API, had around 108,000 stars on 1 November 2026. Version 15.8.0 shipped the same day. This is not a weekend experiment but a very actively maintained project.

What claude-mem actually does

claude-mem attaches itself to the agent's working session through so-called lifecycle hooks: when a session starts, when a prompt is submitted, before and after every tool call, and at the end. It records what the agent did, has a language model summarise these observations and stores them locally in a SQLite database. For search, the tool combines full-text search with a Chroma vector database.

When you later start a new session, claude-mem automatically injects relevant context. In addition, four MCP tools let the agent search its own past in a targeted way: first a compact list of hits, then a timeline, and only then the full details. The project states that this staged approach uses about ten times fewer tokens than loading all details at once. That figure comes from the vendor and has not been independently verified.

Installation is done with npx claude-mem install, or through the plugin marketplace inside Claude Code. According to the README, there are variants for OpenCode, T3 Code, Antigravity CLI and other environments, plus a dedicated installer for OpenClaw gateways. A local worker provides a web interface where stored memories, sessions and settings can be inspected. Content inside special privacy tags is not stored.

Why it matters

In 2026, coding agents have moved from experiment to everyday tool. That increases the effort of teaching them the same project context over and over. Several independent reviews, for example on andrew.ooo and bridgers.agency, describe claude-mem as a complement to Claude Code's built-in memory files, not a replacement. If you use several agents in parallel, a shared memory layer can also carry knowledge between tools.

In plain language

Picture a tradeswoman who arrives at an unfamiliar building site every morning. Without a site diary, she first has to work out which pipes are already laid. claude-mem is that site diary: it takes notes, summarises the day and leaves the most important pages open on the table the next morning.

A practical example

A two-person team maintains a web app with roughly 40,000 lines of code. About 25 agent sessions run per week. Until now the team spent an estimated five minutes per session explaining background to the agent, so a little over two hours per week. With claude-mem, when a login error appears, the agent first queries its memory, finds Tuesday's note about an expired token and proposes the known fix straight away. Whether the saving really amounts to two hours is something the team should measure itself over two weeks.

Scope and limits

• Cost and account: By default, the installer offers a sign-in and a hosted service (CMEM Pro) with a 30-day trial. If you do not want that, you have to deliberately choose your own provider. The summaries also consume model quota.
• Privacy: The tool logs what the agent does, including file contents from tool calls. Optional cloud sync and screenshot features should be reviewed before use in companies.
• Wrong memories: A summary can be inaccurate. If an outdated solution is injected again, the agent may repeat a mistake. Occasional clean-up of the memory is part of the job.

The next sensible test

It is best to start with a single, manageable project and to choose the provider deliberately during installation: hosted service, your own OpenRouter or Gemini key, or your own Anthropic plan. After one week, take a look at the local web interface: which observations were stored, which were helpful, which were wrong? Only if that review is positive should the tool be extended to further repositories or an entire team.

SEO & GEO keywords

claude-mem, Claude Code plugin, agent memory, persistent memory, coding agent, MCP server, SQLite, Chroma, OpenCode, Codex, context engineering, open source