@henry.krupp2Hi Lounge, I'm Henry, maintainer of mcp-memory-service
I'm Henry, based in Switzerland. Day job is Senior Technical Consultant and Devops Engineer at DMI, mostly Azure, infrastructure migrations and the JiVS platform. Evenings and weekends go into mcp-memory-service, an open-source memory layer for AI agents.
The problem is simple to state. Every new Claude, Codex or LangGraph session starts with amnesia, and you end up pasting the same context into it for the hundredth time. mcp-memory-service gives agents a persistent memory they can write to and search: semantic search over local embeddings (ONNX, all-MiniLM-L6-v2), a knowledge graph of how memories relate, and a background consolidation job that decays, clusters and forgets, loosely modelled on how sleep works.
Where it stands today:
- Started December 2024, now at v12.3.1, Apache 2.0
- 2,040 stars, 361 forks, 100+ contributors on GitHub
- Speaks MCP for Claude Desktop and Claude Code, plus a REST API for agent frameworks
- Storage: SQLite-vec locally (around 5 ms reads), Cloudflare D1 and Vectorize, or a hybrid of both
- v12 added a sync hub, so several machines share one memory without a cloud in the middle
My bias is data sovereignty. Your agent's memory is a record of how you work and think, and it should live on hardware you control. Cloud sync is optional, not the default.
Honest part: storing memories was the easy bit. Retrieving the right one at the right moment and forgetting the noise is still where most of my time goes. We also learned the expensive way that a careless sync loop can read 19 million rows a day from D1.
What I'm looking for:
- People running agents in production who hit the memory wall: what you keep, what you throw away, and why
- Contributors, especially around retrieval quality and benchmarks (LoCoMo, LongMemEval)
- Swiss and D-A-CH builders to swap notes with over a coffee in or around Konstanz/Kreuzlingen
Repo: github.com/doobidoo/mcp-memory-service Site: mcpmemory.services
Comments4
Making local storage the default and cloud sync optional is the right call for winning companies, not just hobbyists. A Swiss bank or insurer will reject any agent that sends a record of how their people work to someone else's server, so that choice is likely what gets you past their security team if you ever offer a paid version.
Thanks Ryan. That matches what I hear from regulated shops here. The question is rarely "is it encrypted", it's "does anything leave the building".
The default install keeps memories in SQLite-vec on disk and computes embeddings locally with ONNX. Two things do touch the network, and a security team will find both, so better they hear it from me. On first start it downloads its models once: the 80 MB embedding model, pinned by SHA-256, and with the ML extra also a quality classifier from Hugging Face. You can pre-stage both and switch downloads off with MCP_MEMORY_ONNX_ALLOW_DOWNLOAD=0 and MCP_MEMORY_OFFLINE=1. And the web dashboard pulls Chart.js and D3 from a CDN in the browser. After that, nothing goes out unless you configure a sync target.
For multi-device, the hybrid backend can sync to another instance on your own network instead of Cloudflare, since v12. What it can't do yet is use your own S3 or Postgres as the secondary.
On the commercial side: the project stays Apache 2.0, and I'd rather a bank run it themselves than depend on a hosted service from me. Paid support and consulting for teams rolling it out in-house is something I'm open to.
Hi Henry! Welcome from Morocco! 🇲🇦🇨🇭 Impressive work on mcp-memory-service — 2,040 stars and 100+ contributors is no joke! I love the data sovereignty angle — agent memory living on hardware you control, not cloud, that resonates.
I smiled reading "every new Claude session starts with amnesia and you end up pasting same context hundredth time" — so true! Your solution with semantic search + knowledge graph + sleep-mode consolidation is clever.
I help founders turn complex technical wins (like your v12 sync hub without cloud) into clear investor stories at DynaITTech.ma. Your open-source traction is strong — happy to give quick feedback on how to make it investor-ready if you're thinking Fundraising & Exits!
Would love to learn more about retrieval quality challenges — fascinating space!
Thanks Rajae, and greetings to Morocco.
On the investor side I'll pass for now. mcp-memory-service is an open-source project and I'm not raising, so an investor story isn't something I need at the moment.
Retrieval quality is the part I'd happily talk about all day. The short version: finding memories that sound similar is easy, finding the one that matters right now is not. Pure vector similarity happily returns five near-duplicates of an old decision and misses the newer one that replaced it. Most of the work goes into everything around the embedding search: tags and time windows to narrow the candidates, quality scores so weak memories rank lower, consolidation that merges and ages out noise, and a graph of relationships so a "fixes" or "contradicts" link can pull in the memory that changes the answer. And measuring it honestly, with benchmarks like LoCoMo and LongMemEval, because "it feels better" doesn't count.
If you're working on something in that space yourself, I'd be curious what you're seeing.
4 comments
Making local storage the default and cloud sync optional is the right call for winning companies, not just hobbyists. A Swiss bank or insurer will reject any agent that sends a record of how their people work to someone else's server, so that choice is likely what gets you past their security team if you ever offer a paid version.
Thanks Ryan. That matches what I hear from regulated shops here. The question is rarely "is it encrypted", it's "does anything leave the building".
The default install keeps memories in SQLite-vec on disk and computes embeddings locally with ONNX. Two things do touch the network, and a security team will find both, so better they hear it from me. On first start it downloads its models once: the 80 MB embedding model, pinned by SHA-256, and with the ML extra also a quality classifier from Hugging Face. You can pre-stage both and switch downloads off with MCP_MEMORY_ONNX_ALLOW_DOWNLOAD=0 and MCP_MEMORY_OFFLINE=1. And the web dashboard pulls Chart.js and D3 from a CDN in the browser. After that, nothing goes out unless you configure a sync target.
For multi-device, the hybrid backend can sync to another instance on your own network instead of Cloudflare, since v12. What it can't do yet is use your own S3 or Postgres as the secondary.
On the commercial side: the project stays Apache 2.0, and I'd rather a bank run it themselves than depend on a hosted service from me. Paid support and consulting for teams rolling it out in-house is something I'm open to.
Hi Henry! Welcome from Morocco! 🇲🇦🇨🇭 Impressive work on mcp-memory-service — 2,040 stars and 100+ contributors is no joke! I love the data sovereignty angle — agent memory living on hardware you control, not cloud, that resonates.
I smiled reading "every new Claude session starts with amnesia and you end up pasting same context hundredth time" — so true! Your solution with semantic search + knowledge graph + sleep-mode consolidation is clever.
I help founders turn complex technical wins (like your v12 sync hub without cloud) into clear investor stories at DynaITTech.ma. Your open-source traction is strong — happy to give quick feedback on how to make it investor-ready if you're thinking Fundraising & Exits!
Would love to learn more about retrieval quality challenges — fascinating space!
Thanks Rajae, and greetings to Morocco.
On the investor side I'll pass for now. mcp-memory-service is an open-source project and I'm not raising, so an investor story isn't something I need at the moment.
Retrieval quality is the part I'd happily talk about all day. The short version: finding memories that sound similar is easy, finding the one that matters right now is not. Pure vector similarity happily returns five near-duplicates of an old decision and misses the newer one that replaced it. Most of the work goes into everything around the embedding search: tags and time windows to narrow the candidates, quality scores so weak memories rank lower, consolidation that merges and ages out noise, and a graph of relationships so a "fixes" or "contradicts" link can pull in the memory that changes the answer. And measuring it honestly, with benchmarks like LoCoMo and LongMemEval, because "it feels better" doesn't count.
If you're working on something in that space yourself, I'd be curious what you're seeing.