Picture this: Your company's customer support agent discovers that a major client prefers integrating via API rather than mucking around with the UI. Smart. Three days pass. A sales engineer's agent circles back to that same customer and... starts asking basic setup questions all over again. Meanwhile, in engineering, one developer's coding agent cracks a stubborn Android build issue on Tuesday. By Friday, a colleague's agent is hunting down the exact same bug.
It's a pattern that engineering teams keep hitting once they move past single-agent experiments. Each agent performs brilliantly in isolation. Put them together across departments, though, and they're collectively amnesiac.
Enter Glen, a startup from Y Combinator's Summer 2026 batch that's betting on what founder Nikos Dritsakos describes as "shared memory that becomes shared expertise." The company announced its platform June 1st with a solution designed to solve what might be the defining infrastructure problem of the multi-agent era: how do you make organizational knowledge stick?
A Single Memory Layer for Every Agent
Glen's approach is deceptively simple. Instead of letting each AI agent maintain its own memory silo—or, worse, forget everything between sessions—the company provides a unified memory layer that sits behind every agent in an organization. When one agent learns something useful, every other agent inherits that knowledge immediately. And since agents increasingly mediate how people work, humans get access to that accumulated context too.
The technical implementation leans on the Model Context Protocol, the open standard that lets AI agents hook into external tools and data. Glen runs as a remote MCP server: organizations get an endpoint (something like https://your-org.glen.app/mcp), authenticate once via OAuth 2.1, and point their MCP-compatible agents at it. Works out of the box with Claude, Cursor, Codex, and custom MCP clients.
What happens under the hood is what Glen calls a "Recall + Remember" loop. An agent makes a request. Glen retrieves just the relevant facts from organizational memory that matter for that specific task. The agent does its work. Then Glen automatically captures new learnable facts from the interaction and writes them back to the shared store. Single round trip.
This isn't raw conversation logs piling up. Glen stores discrete, structured observations—facts like "the mobile team uses this specific Xcode configuration for M-series Macs" or "Account Y's API key expires quarterly and needs this particular renewal dance." The backend is Postgres with row-level security ensuring strict organization isolation. Data moves over TLS; storage encrypted at rest. API keys get hashed with argon2id. Access control operates at the observation level through role-based permissions, meaning sensitive information can be restricted even within the shared pool.
Organizations can flip on "private mode" for specific conversations when needed.
The Compounding Thesis
Dritsakos—whose previous startup SalesBop was acquired by SellWell and rebranded as FliteHouse in 2025—published a thesis in May arguing that most memory products are really just "single-tenant storage layers dressed up as cognition." His bet is that organizational memory should accumulate value as more agents and more people feed into it.
It's the compounding effect that Glen's positioning emphasizes. A support agent that onboards in month six already "knows" the institutional decisions, edge cases, and customer quirks the organization learned in months one through five. Glen's materials describe the system as creating "a permanent, auditable record" of decisions and reasoning chains—essentially a queryable timeline of how and why the company operates the way it does.
The YC listing goes a step further, describing how Glen "reconciles" information scattered across code diffs, pull requests, issues, documentation, and meeting notes to give agents a complete picture rather than isolated fragments. Then it "distills [the work] into skills and offers them back to the next agent, or person, who hits the same task."
Bold claims, perhaps. Whether that distillation actually produces reliable skills—versus surface-level pattern matching—remains an open question that only production deployments will answer.
Security Trade-offs
For IT decision-makers evaluating this infrastructure, the security model matters as much as feature promises. Glen's approach tries to thread a needle: organizational-scope sharing by default, but granular access control where it counts. Observation-level RBAC means you can still wall off sensitive information to specific roles or teams.
Authentication uses OAuth 2.1 rather than long-lived API keys sitting in config files. Organizations stay strictly isolated through database-level row security. According to the company's FAQ (accessed in early July), Glen is running in production but onboarding organizations in waves through a waitlist. No public customer names, adoption metrics, or performance benchmarks yet.
A Market Still Taking Shape
Glen arrives as the infrastructure layer for multi-agent systems is still forming. MCP itself only hit general availability in Visual Studio last September. Since then, adoption's accelerated: Microsoft released a Sentinel MCP server into preview around October or November, and WordPress.com added MCP write capabilities in March, letting agents like Claude and Cursor publish content through natural conversation.
The broader concept of unified context layers emerged across multiple vendors early this year. Kong announced Context Mesh around March, positioning it as governance for how agents consume enterprise context. Atlan published an explainer on "Unified Context Layers" in May. Glean, which had been building an enterprise agent platform since at least mid-2025, launched an MCP Gateway in June to standardize organization-wide MCP usage.
Glen differentiates by focusing specifically on shared memory and skills distillation, rather than broader context delivery or governance frameworks. The company's materials critique what they call the "MCP tool fragmentation" problem: most tools hand an agent a single slice of information—one Slack thread, a lone diff—leaving the agent to stitch together the full picture. Glen aims to do that stitching automatically and capture what was learned along the way.
Open-source alternatives exist, though they mostly operate at different scope. Projects like hmem, Agent Recall, Hive Memory, and Ogham—all discussed on Hacker News this spring—provide memory capabilities via MCP but typically focus on per-agent or per-project memory rather than organization-wide shared context.
What Comes Next
Glen is currently working through its waitlist. No public pricing announced. The YC directory listed a team size of one back in early July, though that figure may already be stale. Dritsakos announced the company in a mid-June LinkedIn post, describing Glen as addressing the fragmentation he'd observed in multi-agent workflows.
For organizations already running agents across support, sales, engineering, and data teams, the value proposition is straightforward enough: stop relearning the same lessons in every conversation, every sprint, every quarter. Whether Glen's particular implementation of shared memory delivers on that promise—and whether organizations are ready to trust a unified memory layer with their institutional knowledge—is the gamble that Dritsakos and his YC batch are making.
The memory problem is real. Whether this is the solution remains to be seen.
