Contents

ai-memory looks promising as long-term memory for passing the state of previous work to the next agent. However, handoff quality depends on summary quality, so you cannot judge its practicality until you try it on your own work.

Key points#

  • The goal is for the next agent to receive the status of completed, failed, and unfinished work, even when you switch agents mid-task.
  • It saves summaries and durable notes, rather than full conversations, for the next session.
  • Search centers on SQLite FTS5, combined with entity recognition, adjacency links, and optional embeddings.
  • Setup requires configuring hooks and an MCP server.

What it solves#

When you switch agents, the new agent knows nothing of the previous conversation. It ends up rereading the same files and repeating the same mistakes. ai-memory aims to fill this gap by recording lifecycle events, such as prompts, tool executions, and session boundaries, through lightweight hooks.

Unlike handwritten instructions such as CLAUDE.md, records are created automatically. When a session ends, decisions, open questions, and next steps are consolidated and saved as Markdown tracked in git. Introductory articles also highlight the ability to share memory across multiple coding agents as a strength.

How the handoff works#

After finishing work in Claude Code, open Codex in the same directory. In the example on the public page, Codex starts with a handoff like this:

handoff.md
## 引き継ぎ(Claude Code)
- 完了: refreshをミドルウェアへ移動
- 失敗: Cookieのみだとモバイルで不具合
- 未完了: /refreshのレート制限

Setup involves starting the server, configuring a bearer token, and then running install-mcp and install-hooks for each agent. You can create or resume a unit of work with ai-memory run. Alongside a Docker image, native binaries are available for macOS, Linux, and Windows (WSL2).

A critical look at the caveats#

A summary is still a summary. If it omits the conditions needed to reproduce the cookie issue or the steps that failed, the next agent will try the same path. External commentary also warns that retaining conclusions while losing context can lead to poor decisions later.

Long-term memory also risks carrying attacker-planted instructions into subsequent sessions. Public information alone does not establish how much protection capture exclusions and repository policies provide.

Evidence is also limited. The public page offers one example, and none of the materials explain whether summaries hold up in sessions with many failures or projects lasting several weeks. Evaluation scores are published, but individual usage records are not visible.

Next step#

Choose a small repository with failing tests and try handing it off from Claude Code to Codex once. Check whether the handoff includes reproduction steps and what has already been tried; that will give you a basis for judgment. Installation instructions are available at aimemory.sh.

References#

Source: https://aimemory.sh/ja/

Contents