thedotmack/claude-mem

The problem

Every new agent session starts blank. You re-explain the architecture, the bug you already diagnosed, and the convention you settled on yesterday, and the agent re-reads the same files to rediscover what you both already knew. Context windows help within a session but not across them. Claude-Mem addresses this by capturing what the agent does during a session, compressing it into observations, and injecting relevant pieces into later sessions.

What's different

Hosts usually offer a single long context or a project-level instruction file, which is manual, static and quickly stale. Broader agent frameworks expect you to build your own memory layer, and generic memory services (Chroma/SQLite-backed stores, hosted memory APIs) are not wired into a coding agent's lifecycle. Claude-Mem sits as a plugin/worker that hooks into several agents, does progressive, tiered retrieval so you can see the token cost of each layer, and indexes locally with SQLite plus embeddings. That hook-level integration plus the token-cost transparency is the real differentiator; the extraction itself is an LLM summarisation pipeline whose quality is the whole product.

Who it's for

Developers working on the same repository across many sessions over weeks, especially with Claude Code or a compatible agent (the README lists OpenCode, Codex, Gemini, Copilot, Grok Bot and others), and teams that want a shared workspace on top of individual memory. Skip it if you use a one-off chat, work in an environment with strict data-egress rules, or already have a memory setup you trust — an auto-injected memory layer is one more thing that can go wrong silently.

Cost of adoption

Node 20 or later; a worker service must run. The default installer pushes you toward a hosted account with a 30-day free trial before falling back to your own Anthropic plan, or your own OpenRouter/Gemini key — the free-tier framing and the long list of integrations are the parts to read carefully. Local-only modes exist but are opt-in, and the Grok Bot path reads chat logs from disk. Apache-2.0, TypeScript, fast-moving (version 15 in under a year and a half). Expect breakage on upgrades and an occasional mismatch between what memory recalls and what the code actually says.

Verdict

The idea is correct and overdue: the bottleneck in agentic coding is continuity, not raw model ability. Interest is hard to call established versus spiking: 102,726 stars but only one day on the daily board (+274), with +8,445 in the month, so most of the total arrived in a burst rather than steady growth. That pattern is typical of a project that hit a zeitgeist moment quickly. The codebase is clearly active, but auto-injected memory alters every prompt behind your back, so trial it on a scratch repository first and inspect what it recalls before trusting it on production work.