Memorable, a Y Combinator Summer 2027 company, announced procedural memory for AI agents in a post on X by co-founder and CTO Advaiyt Sane. The two-person startup says its system turns successful agent runs into reusable graphs, rather than requiring agents to start over on repeat tasks.
The product addresses a practical constraint for agent builders: work completed in one context window may not carry forward to the next. Memorable says it records a session, turns successful runs into workflows, connects those workflows in a graph and retrieves relevant workflows for new tasks.
The company contrasts the approach with vector-search and RAG memory. It says its stored procedures contain steps, preconditions, postconditions and the result that determined whether a run passed, rather than text fragments.
According to the Memorable product site, retrieval uses exact, lexical and semantic matching. The company says recall takes about 60 milliseconds and lists Claude Code, Codex, Cursor, gbrain and QM among compatible tools; custom agents can submit JSON traces.
Memorable also says procedures remain in customer-controlled storage, including local files, gbrain databases and Postgres instances. It says only prompts and allow-listed tool arguments leave the machine, while session IDs, user emails, working directories and file contents stay local. Memory use is opt-in and fail-closed, the company says.
The startup reported several customer benchmarks that were not independently verified. It said OpenHome's task-completion rate rose to 100% from 70%; gstack used 293 tokens per prompt versus 15,593, a 98% reduction; and Quartermaster made 40% fewer tool calls, with a 91% pass rate compared with 80% without memory.
Memorable was founded in 2026 and is based in the San Francisco Bay Area, according to its Y Combinator company profile. Its founders are CEO Nikhil Krishnaswamy, a Stanford CS graduate, and Sane, a Carnegie Mellon CS student. Krishnaswamy previously built an iOS Model Context Protocol tool at AGI Inc, while Sane previously built Nvidia developer tools that generate roughly 50,000 autonomous-vehicle tests per month.
The company announced the news in a post on X.
