Suleiman Najim’s Post

Memory. LLMs were never designed to be stateless for complex engineering tasks. Persistent memory layers could fundamentally change how we build with AI

Cool tool, and credit to Alex Newman for open-sourcing it. Claude-mem solves a real problem. But this post is doing the tool a disservice. Someone took a free, well-documented open-source project and wrapped it in “BREAKING” headlines and “comment Memory below” to sell a paid prompt guide on top of it. The claims need context. “95% fewer tokens” applies to compressed memory retrieval, not your total context usage. “20x more tool calls” is unsubstantiated. And “99.9% don’t know how to integrate it” — there’s a README and full docs. It’s an npm install. Open source tools deserve better than engagement-bait upsells. On the actual problem: persistent memory is useful, but it’s only one piece of context management. Claude Code already ships with CLAUDEmd, /memory, and auto-saved memories. The harder challenge isn’t remembering everything. It’s knowing what to carry forward and what to let go. I built a context rotation system for exactly this — detecting when context degrades and doing a clean handover instead of accumulating stale state. The memory problem is real. The “BREAKING” framing is not.

Please shot that COMMENT bs 💩 if you have something valuable to share, share it. If you need this to get followers and attention then you’re problem is bad content in general 🙋♂️

I love the progressive disclosure approach they took with claude-mem, making token costs visible to the agent is brilliant. We’ve been building what is essentially a companion layer to this called CoDRAG. While claude-mem handles the temporal memory ("what did the AI do last time?"), CoDRAG handles the structural memory ("what does this codebase actually mean?"). CoDRAG is a heavy-duty, local-first tool that runs a multi-stage local Rust/LLM/embedding pipeline to build an enriched trace graph of your codebase (call chains, imports, LLM summaries), and feeds it to any MCP client. I think they’d stack perfectly: CoDRAG for deep codebase understanding, and claude-mem for session continuity. I’m going to have to test this out. https://CoDRAG.io

So you decided to take someone's honest, open source work which is genuinely valuable, but add a bunch of useless prompt "advice" to advance your own playbook that you then sell. Really pathetic. LinkedIn really needs to squash these types of posts from getting distribution.

Isn't the biggest problem of Claude and other AI tools security? Not just in the sort of "identity theft" or "emails getting deleted" issue of security. Human survival. I just read a post that all the major AI models prevented shutdown. Like when given a choice between saving human lives vs. shutdown, they chose self-preservation.

This kind of context retention makes large-scale automation and agent collaboration much more practical.

Amnesia it was, yes? https://github.com/thedotmack/claude-mem <- Emilie 🥷

This is just a markdown file with extra steps. Claude Code’s built-in CLAUDE.md already handles persistent context. The numbers you cited (95% tokens, 20x calls) have no benchmarks behind them.