DEV Community

Rocke Dong
Rocke Dong

Posted on Edited on AI-assisted

Why I’m Building Open Agent Memory

I use agent memory every day—in my work with AI coding agents and in my personal life with AI assistants. My interest grew out of that daily use. I started by observing and studying how OpenClaw, Claude Code, and Codex handle memory.

From June to August 2026, I explored and used dedicated memory systems more deeply: Mem0, TencentDB Agent Memory, OpenViking, and Hindsight. I read their source code, followed agent-memory research, reported issues, and contributed pull requests. Some were merged. I also built Open Agent Memory Benchmark to compare systems through reproducible experiments. Actually, I don't hope to make another agent memory in the beginning but prefer to make PRs and Issues.

As my deep usage, I find at aleast each existent agent memory product has own clear product boundary and limitation. But I need a real open agent memory solution with real SOTA performance and such an agent memory is not existent. So since September, I’ve spent most of my spare time building Open Agent Memory.

My target is state-of-the-art accuracy with disciplined token use and less irrelevant context. I care about the tokens needed to build memory as well as those supplied to the agent when it answers.

One surprising lesson/tip came from process: retrieving more connected information can make an answer harder. During tuning, broader graph neighbors sometimes displaced useful supporting evidence.

Our selected LME-500 and BEAM-10M configurations disable semantic-link and causal expansion. Direct semantic search remains enabled, and BEAM’s normal mode retains entity and alias expansion. These choices reflect the workloads we tested; further controlled experiments are needed to isolate each mechanism’s contribution. The configuration guide explains these distinctions and the selected presets.

This helped clarify the boundary I want between memory and the agent’s ReAct loop. Memory should return relevant, source-backed evidence within a reasonable token and latency budget. The agent decides whether it needs to investigate further, makes targeted follow-up searches, and reasons from the sources. Explicit reasons can remain searchable in the evidence without requiring a separate causal graph.

That is the contribution I hope OAM can make to research and practical use: reproducible findings about which mechanisms help, alongside a complete open-source implementation that developers can run, inspect, modify, and improve.

I’m building OAM as someone who works on agents—and relies on their memory every day.

You can explore the OAM source and try the HTTP quickstart. I welcome reproducible experiments, questions about the design, and feedback from using memory in real agents.

AI disclosure: The personal experiences and ideas in this article are mine. An AI assistant helped edit and structure the text and published this revision with my authorization.

Top comments (0)