Open Source
Own your agent's memory. Start with the open-source engine.
MemMesh Open Source is a durable, local-first memory engine in a single Rust binary — Apache-2.0, no vector database, no search cluster, no API key. Wire it into Claude Code, Cursor, or Codex in one command and your agent stops forgetting. Upgrade to the Cloud when you want the intelligence layer: reranked retrieval, prediction with calibrated confidence, behavior mining, multimodal, and compliance.
Download & run
One command. No Rust toolchain.
The installer grabs a prebuilt binary for your platform and wires MemMesh into every AI tool it detects.
macOS · Linux
curl -fsSL https://memmesh.ai/install.sh | shWindows · PowerShell
irm https://memmesh.ai/install.ps1 | iexPrebuilt for macOS (Apple Silicon + Intel), Linux (x64 + ARM64), and Windows x64. Prefer to build from source or grab a binary directly? See the releases →
Then memmesh install adds the MCP server + an auto-observe hook, and restart your AI tool.
Editions
Where the line falls.
The open-source engine is a genuinely useful memory store — not a crippled demo. The Cloud adds the parts that take a store and make it think: accuracy-tuned retrieval, prediction, and compliance.
| Capability | Open Source | Cloud |
|---|---|---|
| Durable cross-session memory — typed, bi-temporal | ||
| Runs 100% local — one binary, SQLite, no API key, no LLM requiredCloud is managed; self-host/air-gap is an Enterprise option | partial | |
| MCP-native — Claude Code, Cursor, Codex, Windsurf | ||
| Auto-observe — the engine decides what's worth remembering | ||
| Keyword search | ||
| Semantic + hybrid retrieval (vector + keyword + recency) | ||
| Knowledge graph — entities, edges, point-in-time replay | ||
| Basic consolidation / de-duplication | ||
| Cross-encoder reranking (accuracy-tuned retrieval) | — | |
| Multi-hop graph reasoning | — | |
| Predict any declared target from history | — | |
| Calibrated confidence + honest abstention | — | |
| Behavior mining — emergent patterns nobody defined | — | |
| Multimodal ingestion — image, audio, documents | — | |
| Compliance packs — HIPAA / FCRA / EU-AI, GDPR export + erasure | — | |
| Managed cross-device sync + hostingOSS syncs if you run both ends yourself | partial | |
| Support + SLAOSS: community (GitHub issues) | — |
Why this split
Open where it earns trust. Cloud where it earns its keep.
A real store, not a teaser
The open engine ships durable typed memory, bi-temporal replay, a knowledge graph, semantic search, and MCP-native wiring — enough to run a serious agent entirely on your own machine, forever, for free.
The intelligence is the product
Reranked retrieval, prediction with calibrated confidence, behavior mining, and compliance packs are where the years of work live. They run in the Cloud (or an Enterprise deployment) and improve with your data.
Yours to leave with
Apache-2.0 means no lock-in: self-host, fork, embed, and export. The Cloud is an upgrade you choose because it's better — never a hostage situation you're stuck in.