Every company eventually develops amnesia. Not the catastrophic kind—just the slow accumulation of details that slip through the cracks. Customer promises buried three layers deep in a Slack thread. Pricing exceptions that live exclusively in someone's head. The bug raised in yesterday's standup, already swallowed by the infinite scroll.
Shepherd, a three-person team that emerged from Y Combinator, believes the fix isn't another wiki or a more sophisticated search bar. It's a living memory system that runs quietly in the background, ingesting everything your company does, then surfaces what you need via text message.
The San Francisco startup, founded in 2026, is positioned with a pitch so direct it borders on blunt: "Your company memory, one text away." Behind that promise is software designed to auto-ingest messages, calls, code commits, and documents into what Shepherd calls a unified memory layer—accessible to both humans and the AI agents increasingly woven into daily work.
Whether that's compelling infrastructure or an overpromise remains to be seen. But the company is betting that convenience—the ability to text your corporate brain from your phone—matters more than yet another dashboard.
Memory That Runs on Autopilot
Shepherd operates through a macOS app and web interface. Once a team connects it, the platform pulls data from tools already embedded in the workflow: Slack conversations, Apple Messages, meeting notes, Notion pages, Gmail, Google Drive, GitHub commits, Linear tickets, Stripe transactions.
The interface demo shows the kinds of queries it's meant to handle. Ask "what did we promise Acme?" and the system returns something like: "SSO plus CSV export by Q3—from the March 4 call and the signed contract." Want to know what a teammate shipped yesterday? Shepherd ties together their code commits with fragments from sales calls.
More intriguing than reactive queries, though, is the proactive layer. Shepherd is designed to ping users first—surfacing follow-ups from partnership conversations, reminding engineers about bug fixes mentioned in passing, flagging tickets that need attention before they go stale.
For AI agents, Shepherd uses Model Context Protocol (MCP) for integrations with tools like Codex, Claude Code, and Cursor as part of the broader MCP ecosystem. The logic is simple enough: give agents the same institutional knowledge humans can query, so they write code that accounts for pricing exceptions or customer commitments without developers having to remember—and re-explain—every time.
The Team
The founding trio brings an unusual mix. Philip Meng, the CEO, studied electrical engineering and computer science at Stanford and previously worked at Animoca Brands and 645 Ventures. Ishan Ramrakhiani, the chief product officer, comes from Stanford's biomedical computation program with research roots at Dartmouth. Elijah Renner, the CTO, is finishing high school while beginning at Stanford for computer science—he's already logged medical NLP research at Stanford AIMI and Dartmouth-Hitchcock.
It's an odd configuration: a team building enterprise memory infrastructure while one founder is still technically a high schooler. Then again, YC's Diana Hu, the startup's primary partner, has backed stranger bets. Sometimes the odd configurations are the ones that stick.
A Market That's Anything But Lonely
Shepherd is walking into what might charitably be described as a crowded room. The "company brain for AI agents" category has seen an explosion of entrants recently, with at least a dozen direct competitors visible.
There's Hyper, also out of Y Combinator, positioning itself as "the company brain that powers your AI employees." Monora describes itself as a "sharing layer for your company brain" with agent compatibility through MCP. Wemory talks about "compounding collective intelligence for AI-native teams." AIOS offers "agent workspaces and a shared team brain." Sentra, Teamkit, Inherent, and /.relay all stake claims to variants of unified memory or brain-like architecture.
The enterprise incumbents aren't standing still either. Glean has added native agent memory and launched multiple agent-focused products recently. Notion rolled out AI meeting notes with messaging around giving teams "perfect memory," alongside agents capable of executing multi-step actions.
The differentiation game here is subtle—sometimes maddeningly so. Some platforms emphasize on-premises deployment. Others tout governance frameworks or citation-auditable answers. A few focus on portability: Walrus Memory, for instance, announced itself as a "portable memory layer for AI agents," stressing verifiable and transferable knowledge.
Shepherd's angle appears to be access simplicity: the ability to text your company memory from your phone, combined with deep MCP integrations for developer workflows. Whether that carves out durable differentiation is very much the open question.
The Trust Problem
According to Shepherd's terms of service and privacy policy, the platform maintains detailed controls over what it ingests and how. For coding sessions, a local collector reads session records from Codex and Claude Code, redacts sensitive values locally, then uploads repository metadata plus structured messages—but not raw tool outputs.
The company states it adheres to Google's API Services User Data Policy and doesn't use ingested data for advertising or to train generalized models. Human employee access is limited to service, security, legal, or support cases. Some deployments may allow stricter configurations, though the default posture isn't zero-access by design.
Still. Handing over your entire company's communication history to a three-person startup requires a leap of faith. As of July 3, 2026, Shepherd doesn't yet list customer logos or case studies on its website—the classic early-stage paradox. There's no public pricing visible either, just a "Talk to a founder" call-to-action, which suggests they're still figuring out the packaging.
Infrastructure, Not Magic
Y Combinator's standard deal for the S26 batch involves a $125,000 investment for 7% post-money SAFE and $375,000 via an uncapped most-favored-nation SAFE, totaling $500,000 in Shepherd. The startup was also reportedly in line for Sam Altman's offer of $2 million in OpenAI credits to companies in the YC batch, structured as an uncapped SAFE that converts on the next priced round.
The macOS app and web interface are live. The integrations work. The market is absurdly crowded, which either signals validation or saturation depending on your read of the moment.
For now, Shepherd is a bet that company memory infrastructure becomes essential in an agent-heavy world—and that the team who makes it easiest to access, whether via text message, MCP, or whatever interface feels natural, might be the one that endures.
The harder question is whether "easiest" is enough when everyone else is building roughly the same thing. That answer won't arrive via text message. It'll take time, traction, and probably a few pivots along the way.
