Ahmed Lasheen's 28-Tool Memory Stack Breaks Claude and ChatGPT Lock-In
160 Hacker News points later, someone finally built the memory layer that every AI developer has been secretly wanting. Ahmed Lasheen's Stash project does something beautifully subversive: it gives any AI agent the same persistent memory capabilities that Claude and ChatGPT hoard behind their paywalls.
Here's the kicker. While everyone was obsessing over which model has better reasoning, Lasheen focused on the real competitive moat: memory. Claude's Team/Enterprise accounts just rolled out memory summaries in September 2025. ChatGPT auto-injects 33 long-term user facts plus summaries from your last 15 chats. Both companies want you locked into their ecosystem.
Stash says: not so fast.
The Reverse-Engineering Revelation
Developer Mantan Gupta's September analysis revealed something fascinating about how these systems actually work. Claude is surprisingly transparent - it shows you exactly when it's searching your conversation history through visible tool calls like conversation_search and recent_chats. No AI summaries, just raw keyword searches.
ChatGPT? Total black box. It silently layers system instructions, session metadata, user facts, and chat summaries before you even start typing.
<> Simon Willison noted Claude's approach is "token-efficient but model-dependent" while calling out the transparency advantage over ChatGPT's "opaque auto-injection."/>
This architectural difference matters more than most people realize. Claude's method is hackable. ChatGPT's method is... well, ChatGPT's method.
What Nobody Is Talking About
The timing here is suspect in the best possible way. Three similar projects dropped almost simultaneously:
- Stash (Apache 2.0, 28 tools, MCP server)
- Mem0 Chrome Extension (cross-LLM memory with one-click sync)
- OGHunt's AI Memory Layer (30-second setup for MCP-enabled models)
This isn't coincidence. This is a coordinated assault on memory lock-in. The developer community clearly reached a breaking point with proprietary memory systems that trap your conversation context inside one company's walls.
Stash's 28 tools and background consolidation features suggest Lasheen didn't just clone existing functionality - he improved it. The MCP server integration means any compatible model can plug in immediately.
The Token Economics Nobody Mentions
Here's what makes this technically interesting beyond the obvious "screw Big Tech" angle. Traditional RAG and vector databases are apparently overkill for memory.
Both Claude and ChatGPT ditched the fancy stuff for simpler methods:
- Claude: Direct keyword search through raw conversations
- ChatGPT: Layered context injection with sliding windows
- Stash: Knowledge graph storage with intelligent consolidation
The last approach might actually be the smartest. Knowledge graphs prevent the token bloat that comes from injecting full conversation histories, while avoiding the context-missing risks of Claude's model-dependent recall triggers.
The Real Victory Condition
Lasheen's Apache 2.0 licensing choice reveals the end game. This isn't about building a sustainable business around memory-as-a-service. This is about commoditizing memory so thoroughly that no single company can use it as a competitive moat.
The 70 Hacker News comments were mostly positive, which tells you everything. When developers see a tool that "brings the same capability to any agent" and liberates them from platform lock-in, they don't nitpick the implementation details.
Will this actually threaten Anthropic and OpenAI's memory strategies? Probably not directly - most enterprise customers will stick with integrated solutions. But it absolutely changes the game for indie developers, small firms, and anyone building custom agents.
The memory wars just got interesting. And messy. Exactly how we like it.
