Should you try Claude-Mem?
Try it when automatically recorded coding history would help you resume work and you can validate the complete capture path yourself. Claude-Mem captures tool-use observations, generates summaries, and makes them available to later sessions. Its documented search workflow retrieves an index, adds chronological context, and fetches selected observations. Our sandbox checked repository setup commands; it did not establish whether those memories are accurate or useful in practice.
What our sandbox established
Our run was recorded on October 6, 2026, at commit 115540c. Installation completed in 141 seconds, and the build completed in 4 seconds. These are command outcomes from our sandbox, not timings for the end-user plugin installer.
The harness records each step's success flag, exit code, elapsed time, and timeout status, alongside a retained log tail. The test step returned exit 127 after 1 second. Its script invoked bun test tests, and the shell reported bun: not found. No assertion failure or passing-test count was established. The build tail also says Bun lockfile regeneration was skipped because Bun was unavailable, despite the overall build returning zero.
The build output identifies the checkout as version 13.31.0. The separately supplied v13.34.2 release notes describe a later release, not the version exercised by our run. We did not start a live coding session, connect an observer provider, verify worker recovery, or measure recall quality. A successful build does not establish any of those outcomes.
The testing-methodology page explains how MrKeyoor reports command results and their limits. Its supplied October 31 snapshot provides provenance context; it does not establish which procedure version was used on October 6. This correction uses the existing stored run and includes no new execution.
Read the stored summaries alongside the raw output
The attached dated harness values are stored parser summaries: 800 dependency packages, a 1,033 MB installation footprint, and an npm audit total of 0 recorded advisories. The raw install tail first shows an npm ci lockfile error, then a completed installation with different package figures and nonzero advisory output. The install step's final recorded status is successful. The supplied record does not reconcile the differences between that tail and the stored summaries.
We retain the stored measurements and disclose the disagreement rather than treating either output as an exhaustive dependency inventory. Audit scan coverage is unknown because the record has no scannedPackages field. The recorded zero therefore supports neither a package-coverage claim nor a security verdict. Test counts are absent, and parser summaries alone would not establish exhaustive unique test coverage.
What you would operate
The README describes a Bun-managed local HTTP worker, SQLite storage for sessions and observations, and Chroma for hybrid semantic and keyword search. Its staged retrieval design lets an agent inspect candidate memories before fetching their full contents. We did not measure the README's token-saving claims.
The README also distinguishes installing the SDK globally from registering plugin hooks and setting up the worker. Use the documented plugin installation path for your host, then verify that a harmless observation is captured and can be found in a later session. Inspect what was stored, check exclusion controls, and choose an observer provider appropriate for the material being processed. Account sign-in can be skipped through the documented provider or opt-in settings.
The Codex report is closed, with a recovery limitation
Issue #4562 is closed. The maintainer says the adapter fix shipped in v13.29.0 and restored the tool-use backup rows missing in the reporter's v13.28.0 setup. That report should not be presented as an unresolved defect in every current installation.
The same response says accepted events still wait in memory until summarized, so a restart can drop pending events; the backup row provides a durable copy. Test restart recovery on your chosen version before depending on it. We did not verify the fix or recovery behavior in our sandbox. The supplied independent extracts do not establish the Chroma issue's current status, so this review makes no current-defect claim about it.
The v13.34.2 release notes describe a prompt-cache regression fix in the shared OpenAI-compatible observer loop. They explicitly say production provider-cost effects were not measured. Their validation results belong to the project's release report and are not our sandbox's passing-test counts.
Choose a framework when you are building the memory system
Graphiti is a framework for building and querying temporal context graphs inside your own application. Its README says you supply the graph database, model, embedder, and surrounding service. Choose it when changing relationships and historical facts are central to the application you are building. It is an alternative architectural approach, not an established drop-in replacement for Claude-Mem's coding-session capture.
LangMem provides memory primitives that work with any storage system and native integration with LangGraph's storage layer. Consider it when you want to implement memory behavior within your own agent application. Its example uses an in-process store that loses memories on restart; the README directs production users toward database-backed persistence. Neither framework was executed in our sandbox, and this review establishes no ready-made integration with Claude-Mem.
How to read the ratings
These are editorial assessments, not benchmarks. Setup ease runs from 1 for blocked setup to 5 for verified operation with little intervention; 3 reflects successful installation and build with unresolved runtime setup. Documentation runs from 1 for missing instructions to 5 for instructions validated end to end; 4 reflects substantial installation, configuration, and architecture material without that full validation.
Community rates visible public support channels and response evidence, not adoption or popularity: 1 means neither is evidenced, 3 means channels exist, and 5 requires a substantive maintainer response as well. Maturity rates documented release and recovery practices: 1 means no evidence, 3 means documented releases with substantial operational unknowns, and 5 requires independently verified operation and recovery. The maturity score of 4 recognizes a documented regression fix and explicit limits while reserving the highest rating for verification we have not performed.