What annoys you most about opening Claude Code to write code?
The agent doesn't remember you said "no TypeScript in this project" last week. Doesn't remember you prefer functional style. Doesn't remember you rejected an architecture proposal last time.
You have to say it all again. Every single time.
agentmemory wants to solve exactly this.
8,483 stars, up 4,450 in a week, 301 commits, supports Claude Code and Codex plugins.
How It Works
agentmemory's core idea is simple: persist the information the agent gains during conversations, and auto-load it on the next startup.
But "simple" hides a lot of engineering detail.
Three-tier storage. The project divides memory into three types:
- Project-level memory: architecture decisions, tech stack, code norms for this project. Survives session restarts.
- User-level memory: your coding preferences, common patterns, personal habits. Cross-project.
- Temporary memory: current conversation context, cleared when session ends.
This分层 is reasonable. You don't want the agent to remember every historical conversation—that introduces too much noise. But architecture-level decisions should persist.
Search and retrieval. It doesn't just dump all memory into the context window. The project uses hybrid search with BM25 + vector retrieval, and recently added CJK tokenizer support for Chinese/Japanese/Korean text—important for Chinese projects.
Public benchmarks. There's a benchmark directory with 100k record load tests, publishing p50/p90/p99 latency data. This level of transparency is rare in similar projects.
My Actual Experience
I installed the .claude-plugin into Claude Code, and deployment was fairly smooth.
It works—it did remember I said "no classes, functional style" in one project, and didn't generate class code in the next session.
But there are pitfalls:
- Memory updates are delayed. I changed project preferences, and the agent sometimes still used old memory. Looks like cache refresh has lag.
- Memory overload. After running for a while, the memory store accumulated dozens of entries. The agent started confusing things—it remembered too many details and lost the point.
- Privacy concerns. Memory is stored locally, which is good. But if your project has multiple collaborators, everyone's preferences mixing together becomes a problem.
What Real Problem This Solves
The biggest pain point of agent coding isn't insufficient capability—it's context断裂.
You spend an hour with the agent today, clarifying project architecture and settling on a technical approach. Tomorrow you open a new session, and everything resets.
agentmemory's value is narrowing this gap. It can't make the agent fully "remember" you, but at least it remembers key decisions.
It's Still Far From Done
Don't be fooled by the star count. This project is far from "production-ready."
The biggest issue is automatic memory management. Who decides what to remember and what to forget? Currently the mechanism is basically manual—you tell the agent what to remember, and it remembers. But users can't precisely tell the agent "store this in project memory, store that in temporary memory" every time.
What's needed is auto-classification, auto-cleanup, auto-summarization. That doesn't exist yet.
My Take
agentmemory's direction is correct. Coding agents need persistent memory—that's inevitable.
But if you're relying on it for important decisions right now—don't. Treat it as an auxiliary tool to reduce the burden of repetitive explanation. That's enough.
Once the project nails automatic memory management and cross-collaborator isolation, then seriously consider integrating it into your workflow.
Primary Sources:
- GitHub - rohitg00/agentmemory — 8,483 stars, 301 commits
- GitHub Trending Weekly — 4,450 star growth this week