Claude-Mem

Claude-Mem summarizes observations from coding agents and retrieves relevant history when work resumes in another session.

Design a narrow memory handoff for a coding project

Choose a recurring context gap and use Claude-Mem’s search, timeline and selected records as a model for a lightweight handoff.

Name the project, typical interruption and the source records that may be retained.

Copy the request into a new Floatboat conversation and provide the required material.

Download Floatboat

Your request for Floatboat

Preview and copy the request. It is not sent automatically from this page.

Reference: Claude-Mem
https://github.com/thedotmack/claude-mem

Read the original documentation and check the current environment and dependencies. Propose any new installation or tool connection before proceeding.

My task:
Read the Claude-Mem architecture and search docs. For our checkout bug investigation, design a handoff with observation IDs, one timeline window and a branch/test reference. Define what must stay out of memory and how to verify a recalled claim before editing code.

Expected deliverable:
A decision-log schema, retrieval checks and an explicit retention and review boundary.

How Claude-Mem recovers context across sessions

When a coding session ends, its tool outputs and decisions can be hard to recover. Claude-Mem captures observations through host-specific integrations, creates summaries and stores session records that can be searched later. The local worker and viewer make it possible to inspect which records were saved.

Search starts with a compact list of observation IDs. The timeline tool shows nearby activity, and selected records can then be opened in full. Startup context has its own settings. For Claude Code and Codex, it defaults to observations from the current host, with a setting to include other sources.

A Floatboat handoff can borrow this pattern by storing a small decision log and retrieving only what the next task needs. Compare recalled claims against current code and tests, and decide which material can be recorded before enabling hooks, observation providers or optional sync.

What Claude-Mem keeps and retrieves

The capture, storage and search paths do different jobs. Test the worker with a disposable repository before choosing how much past context to return in a new session.

  • Session observation hooks

    Capture tool activity through the integration for your coding host. Check what that host can observe before choosing where to enable capture.

  • Local worker and viewer

    Inspect saved observations and summaries in the web viewer. The local worker keeps sessions, observations and summaries in SQLite.

  • Layered MCP retrieval

    Search for observation IDs, look at surrounding events, and fetch complete records only for selected results.

Claude-Mem use cases

Search the observation index for a past decision, then retrieve only the details you need. Check the old record against the current branch and tests before reusing the decision.

Resume a bug investigation after a session reset

Search for a symptom, failing test or function name, then read the nearby observations. Match the earlier reproduction steps against the current branch and rerun the failed check before changing the implementation.

Recover the reason behind an old design decision

Look for the decision and the events leading up to it. Compare the retrieved explanation with the commit and any later changes so a past workaround does not become an accidental requirement.

Review context before moving to a different coding agent

Check whether the receiving host should include observations from other sources. Try with a synthetic handoff first and inspect the resulting startup context to make sure it contains the intended project and scope.

Audit what should be remembered from a test session

Run a harmless task with approved sample text. Open the local viewer and compare what was stored with what the workflow was allowed to retain, including a test of the documented private-content marker.

What to check before enabling memory

Choose a supported coding agent and a disposable repository for the first run. Check which events it captures, what a fresh session receives, and whether inference or sync leaves the device.

  • A supported coding host

    Follow the install guide for your host. The Claude Code setup documents Node.js 20 or later and a version of Claude Code with plugin support.

  • An explicit data scope

    Choose what the plugin may record and review the provider and sync settings before running it with customer data or internal repositories.

  • A way to inspect retained context

    Open the local viewer and try a search, a timeline lookup and retrieval by observation ID with your sample session.

Why AI forgets between sessions

Learn which decisions and project records help a new session resume work.

Related Software Projects

Compare the workflows, operating requirements, and source licenses of these projects.

Claude-Mem FAQ

Is Claude-Mem made by Anthropic?
The upstream README credits Alex Newman, whose GitHub handle is thedotmack. Claude Code is a host the project integrates with.
Does it store every full session in the context window?
The project captures observations and summaries. Startup context can be limited, and the MCP flow retrieves the full record only for selected results.
Does cross-host context injection happen by default?
The README documents that Claude Code and Codex startup use the current host’s observations by default. Including all sources requires a setting change.
Is local storage the same as local inference?
No. The worker can keep records locally, while summarization uses the chosen observer provider. Review provider settings and optional sync separately.
What should I check before capturing confidential work?
Inspect what the selected integration captures, test the private-content exclusion in a disposable project, and review the observer and sync destinations.
What does the Apache-2.0 license cover?
The upstream license documentation covers the public core, including the memory engine, server, CLI, SDKs and adapters. Hosted and reserved commercial services are described separately.