One of the very first problems I hit when I started to do agentic programming was the context problem, and that my agent always started again from the beginning. Being a bit naive and not really knowing what I was doing, I started off writing a mcp tool (because those are cool, right?) to solve the problem.

MD files are primitive, I thought. This memory stuff should really be a database, I thought. So I tried a couple of free memory tools, but nothing was really doing anything for my particular itch. I wanted the agent to remember what I needed, when I needed it, and with the minimal of effort. I wanted it to talk architecture with me, file everything away, and be able to remember it when I wanted to write the stories I needed for the software I was building, and faffing around with plans and files was a pain. I need a knowledge graph.

To be honest, the tools I built are not required if you are happy with an md file solution and your claude.md file. These tools aren't essential for you if that's the case. If you have progressed beyond that set up, carry on reading!

The first tool I wrote was memoryweb (https://github.com/corbym/memoryweb, MIT). In a series of conversations with Claude, we worked out the basics of the api from the agent's perspective and had dogfooded a pretty good tool in just a few weeks. I was using it at work almost constantly. When it went wrong, I asked the agent why, and used memoryweb to record the problem and fix it with a story. It was a very effective loop, and almost effortless.

What about a multi-tenanted version? I had loads of important discussions that my team might like to keep track of. Recordari exists mainly because:

* When two users write conflicting decisions to the same file, no one will notice

* Decisions can be superseded, but old text remains, and the agents can think the stale version is the right version

* After multiple sessions, no one can say why a decision was made or if its true

* Nothing gets recorded about when or who made the decision effectively

Recordari was specifically written to deal with those things. Memories are typed, and the agent can connect memories with typed edges that label nodes "contradicts" and "resolved" - exactly the type of information you'd need to know whether someone disagrees. The server flags semantically close memories that could conflict, agents then adjudicate the possible contradictions and resolves them. Every memory has an owner, so agents don't just overwrite human decisions.

Oh right, the numbers. Yeah.. so the numbers were run using Memora's (Microsoft Research, ICML 2026) open-source benchmark harness, unmodified. The same model, the same LoCoMo dataset and the same category-exclusion conventions. Recordari scored 0.831 end to end, the LLM judge accuracy vs Memora's published 0.801. All the results are public (MIT): https://github.com/corbym/locomo-recordari. The biggest surprise was the multi-session questions coming out at 0.869 vs the raw storage baseline of 0.760.

The harness itself runs against the prod API - exactly the MCP search and recall api that a real agent would use. Not a mocked backend, the real API, with a configuration of `top_k=30`, and a 1 hop edge expansion.

Recordari works with Claude, Claude Code, ChatGPT and pretty much any agent that can connect to MCP and reason.

If you want to run an agent against it to run the LoCoMo benchmark yourself, you can anon login at:

https://admin.recordar.io/start

Just remember to add the Full Skill from the Connect page, and then Connect your agent. Et Viola.

What I would like from HN: have you felt the same pain with a team as me? I'd also like to know if this isn't something you...