claude-mem is a persistent-memory plugin for AI coding agents: it captures what an agent does during a session, compresses it, and injects relevant context into later sessions. This claude-mem persistent memory review covers the repository, README, and release notes as of August 19, 2026. It is a source and documentation review, not a runtime test.

Answer in brief: claude-mem is worth a pilot if your team already works daily inside Claude Code or another supported agent and losing context between sessions costs real time. It is not a set-and-forget add-on: the installer now leads with a paid CMEM Pro provider, recent releases change settings behavior, and the project moves fast enough that you should pin versions and read changelogs before upgrading. As of this review, the latest release is v13.15.2, published on August 16, 2026, and the repository reports 91,153 stars.

Key facts at a glance

What claude-mem actually does

According to the maintainers, claude-mem captures tool usage and session activity, generates semantic summaries, and makes them available to future sessions. The README describes five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd; six hook scripts), a local worker service managed by Bun with an HTTP API and web viewer, SQLite storage for sessions, observations, and summaries, and a Chroma vector database for hybrid semantic and keyword search. Search runs through a mem-search skill and MCP tools using a three-layer workflow: search for a compact index, timeline for surrounding context, and get_observations for full details.

The maintainers list Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, and OpenCode as supported targets. That breadth is the strongest reason to look at claude-mem rather than a memory feature tied to one tool. The counterweight is that behavior depends on hook integration per tool, so the experience will not be identical everywhere.

Who should consider it

The realistic fit is a solo developer or small team that lives inside Claude Code or a supported agent on real projects, where context loss between sessions costs real time: resuming a debugging session, keeping a long migration on track, or handing an investigation to a teammate. If your agents already get enough context from project files and written conventions, a memory plugin adds moving parts without much payoff.

Teams that need audited persistence, strict data-residency guarantees, or an SLA around the memory layer should treat claude-mem as a candidate to test, not a given. This review does not verify retention, encryption, or privacy behavior beyond what the README claims.

What changed in v13.15.2

v13.15.2 is mostly an operations release. The maintainers say the memory observer now explains what happened and what to do when it stops working, most commonly an exhausted CMEM Pro allowance, instead of running silent retry loops around OpenRouter 502 errors. The release parses error envelopes from the cmem.ai gateway and classifies 402 and key-limit-exceeded bodies as quota exhaustion, which it does not retry.

Operationally, each observer failure now produces one log line instead of five, the session-start warning shows a message, a What to do line, a link, and a request ID, and the warning appears even for projects with no memories yet. The release also adds observer-health alerting at session start when observations stop flowing, and surfaces a 7-day Pro trial wherever the viewer URL is shown. A related cmem.ai gateway change introduces honest status codes (402, 401, 429, 503, never 502), a six-code error taxonomy, and an x-request-id on every error. The maintainers note that v13.15.1 was tagged but never released; 13.15.2 supersedes it.

For a team already running claude-mem, this release is about knowing why the pipeline failed. For a new adopter, it matters because there is a commercial gateway in the default path and quota exhaustion is an expected failure mode.

Costs and provider choices

Starting with v13.14.0, the installer asks which memory provider to use and shows an estimated cost per 1,000 observations. Per the release notes: CMEM Pro is first in the prompt at a flat $30 per month with cloud sync included and $0 per 1,000 observations; OpenRouter is listed at about $2.73 per 1,000 observations; Gemini at about $3.39; and billing through your Anthropic plan at about $8.91. v13.15.0 then added an end-to-end 7-day Pro trial inside the installer, with device-code approval and pricing fetched from the API rather than hardcoded.

Two consequences follow. First, per-observation cost depends on how chatty your sessions are, so the $30 subscription is the predictable choice or the expensive one depending on volume. Second, the default path now points at a cloud service with sync and a trial funnel, which changes the question from “is this local” to “which parts stay local.” The v13.14.0 notes show the provider base URL and model come from settings, which suggests an OpenAI-compatible path exists, but this review does not verify how complete or supported it is.

What to check before production

• Install correctly: the README warns that npm install -g claude-mem installs the SDK/library only and does not register plugin hooks or start the worker. Use npx claude-mem install or the /plugin commands, then verify the hooks and worker after install.
• Pin versions and read changelogs: releases are coming fast: v13.13.0, v13.13.1, v13.14.0, v13.15.0, and v13.15.2 all landed in August 2026. The README version badge still shows 13.4.0, so documentation lags the code; do not trust the README for current version numbers.
• Review settings migrations: v13.13.0 changed the default CLAUDE_MEM_TELEGRAM_TRIGGER_TYPES to security_alert,sensitive and auto-migrates existing installs whose list is exactly the old default. The maintainers note that if you deliberately set the list to exactly security_alert, the migration cannot tell that from the seeded default, so set something explicitly different after upgrading if you want to opt out of sensitive-type notifications.
• Map the data path: observations live in local SQLite per the README, but CMEM Pro includes cloud sync, and the plugin captures session activity including the sensitive observation type. Confirm what is captured, where it is stored, and whether anything leaves your machine before pointing it at proprietary code.
• Check runtime requirements: Node.js 20 or higher, a current Claude Code with plugin support, Bun and uv auto-installed if missing, and bundled SQLite. Windows users may need Node.js and npm on PATH, per the README setup notes.
• Stay on the stable channel: the maintainers ship stable releases from main to npm; core-dev and community-edge branches run from source. Production installs should use published releases, not source branches.
• Understand failure behavior: v13.15.2 improves observer error reporting and health alerts, but you still need a plan for what happens to your workflow when the observer or provider quota fails. The memory pipeline is a dependency now.

Maintenance signals

The repository was created on August 31, 2025, and reports 91,153 stars and 7,971 forks as of this review, with the last push on August 18, 2026. That is exceptional adoption for a project under a year old, matched by a very active release cadence. The counterweight is 213 open issues, documentation that lags the latest release, and a repository that GitHub metadata puts at roughly 305 MB, a large clone for constrained machines.

There is also a commercial-community layer to be aware of. The README prominently promotes CMEM Pro, and it includes a section describing CMEM, a token created by a third party that the creator officially embraced, with a contract address. That is not a security verdict; it is a signal about project positioning that belongs in any long-term dependency decision.

Limitations and evaluation questions

This review did not install or run claude-mem. Claims such as the roughly 10x token savings from the three-layer MCP workflow are the maintainers’ numbers and need validation in your own workload. Auto-install behavior, hook reliability, and search quality depend on your environment and agent versions.

• What does a typical session capture, and how much of it is noise you would not want in future context?
• What is your real per-observation volume, and which provider price wins at that volume?
• Can the memory layer point at a self-hosted OpenAI-compatible endpoint, and is that path documented and supported?
• Does settings auto-migration change any of your custom configuration after each upgrade?
• What happens to memory access when the observer fails, and does your team know how to diagnose it? v13.15.2 makes this easier, not automatic.

Alternatives worth comparing

The repository’s own topics list mem0, openmemory, and supermemory, so claude-mem is one option in an active memory-tool ecosystem. The baseline alternative is no plugin: written conventions and context files in the repository, which cost nothing and keep all context under version control. Compare claude-mem against that baseline first; if project-file context is not scaling, compare it against the other memory projects on the same criteria: capture quality, provider cost, data location, and upgrade stability.

Verdict

For a team already running Claude Code or another supported agent where session context loss is a measurable problem, claude-mem is a reasonable candidate, on the condition that you run a small pilot with a fixed provider, review the data path, and pin releases. For teams that need a stable, audited persistence layer, want to avoid commercial upsell paths, or run agents only occasionally, the added dependency is not justified by the evidence reviewed here.

This verdict applies to the v13.15.2 snapshot of August 19, 2026. The project changes quickly, so re-check the current release and changelog before deciding.

Sources

• claude-mem repository (README, license, and metadata)
• claude-mem v13.15.2 release notes
• claude-mem v13.14.0 release notes
• claude-mem v13.13.0 release notes
• claude-mem documentation

Where I can help with adoption

If you are evaluating claude-mem for a team or a production workflow, the gap between the README and a working rollout is real: provider and cost modeling, settings and migration review, data-path checks, and a controlled pilot on your actual agent sessions. That is the AI-agent integration work I do at XP812, and a practical next step is a small scoped trial with the evaluation questions above answered in writing before you commit.