Engineers juggling multiple AI assistants face a frustrating pattern: every new session begins cold. Claude Code helps with architecture, Cursor handles refactoring, Codex manages deployment, but none of them remember what happened an hour ago in a different tool. Glen, a Y Combinator-backed startup, thinks it has found a workaround.
The company launched a shared memory layer that captures context from past agent sessions and feeds it back when engineers start fresh work. Glen pulls from Slack channels, GitHub pull requests, support tickets, call transcripts, and documentation, then injects relevant history automatically when an agent spins up. According to an internal benchmark the company published in August—which involved replaying the last 100 merged pull requests in a workspace that had used Glen for two months—agents using Glen completed tasks with 29% fewer tokens and ran 21% faster than those working without it.
Whether that efficiency survives contact with messier real-world codebases remains to be seen. Glen's founder, Nikos Dritsakos, is betting that shared memory becomes infrastructure as coding agents proliferate.
The Cold-Start Problem
Most engineering teams now run several AI agents in parallel, but each session treats prior work as if it never happened. Glen addresses this by building what amounts to a shared notebook that any agent can read from and write to. The system captures transcripts from Claude Code, Codex, and Cursor through an MCP (Model Context Protocol) server, according to the startup's site. It also ingests Slack messages from public channels, GitHub activity, and team documents.
When an engineer fires up a new coding session, Glen feeds relevant context into the agent's prompt without requiring manual setup. The read-and-write process happens in the same call, the company said, so workflows don't stall while the system searches memory.
Glen runs row-level security to isolate organizations and checks access control on each observation at recall time, per the company's documentation. A "private mode" option excludes sensitive sessions from the memory store entirely. The system reads only public Slack channels and shared team transcripts.
For storage and inference, Glen uses Neon for managed Postgres and OpenAI for embeddings and LLM calls, both under API terms that prohibit training on customer data, according to the startup's subprocessors page last updated in early July.
A Solo Founder With One Exit Already
Dritsakos, 24, previously led special projects at Composio, an AI tooling company, where he held a member-of-technical-staff role. Before that he served as VP of Technology at FliteHouse at 22, a position he took after his first venture, SalesBop, was acquired and rebranded as FliteHouse. That startup, which he bootstrapped, ran for 18 months, according to Y Combinator's directory and a 2025 profile in McMaster University's startup publication The Forge.
Glen is deployed organization-wide at Composio, scaling from two initial users to roughly 20, according to the company's self-reported figures in its Y Combinator launch post published around late August. Glen said the system is also live at "over a dozen other companies" in San Francisco. The homepage displays logos for Zaplar, Litmus, Corgi Insurance, Zima Labs, and Amorphic Labs, though the company hasn't published detailed case studies yet.
In the 30 days leading up to its YC launch, Glen said it injected "tens of millions of characters of organizational context" into agents across its user base, a self-reported metric not independently corroborated. The August benchmark involved replaying the last 100 merged pull requests in a workspace that had used Glen for two months. Tasks completed with Codex 5.5 using Glen consumed 29% fewer tokens and finished 21% faster than identical tasks without it, with quality scores at parity or better, the company reported. Glen rolled back memory state before each test to avoid data leakage.
A Crowded Space
Glen is far from alone in chasing shared context for AI tools. Glean, an enterprise search company with significantly more funding, launched independent agents built on its context layer earlier this year, emphasizing unified access to Jira, Slack, and Teams data. Newer entrants include Glia, which offers a local-first AI memory bridge with MCP hooks, and Scritty, positioning itself as searchable shared memory for coding agents, both visible on Product Hunt.
Glen operates through an MCP server, a command-line interface, plugins for Claude Code and Cursor, a Slack bot, a pull-request reviewer, and a web app. The startup hasn't disclosed pricing and currently runs a waitlist, per its homepage as of mid-September.
What Comes Next
Glen is hiring its first technical employee in San Francisco at $165,000 to $225,000 salary plus 0.5% to 3% equity, according to a Y Combinator job posting. According to that posting, the company is willing to sponsor visas and considers new graduates. Right now, the headcount is one: Dritsakos himself.
The bet is straightforward. As AI agents become standard tools in engineering workflows, the pain of context-free sessions grows more acute. If Glen can make memory sharing reliable enough to trust, the inefficiency it eliminates might justify another layer in an already crowded stack. The alternative is engineers continuing to re-explain the same project details to fresh agent instances, burning tokens and time with every new session.
