A New York startup wants to solve a persistent problem in AI development: agents that forget everything between sessions.
Asymptote Labs announced Tuesday it has raised $5.5 million in seed funding to build commercial infrastructure around Beacon, its open-source platform that gives coding agents something they currently lack—memory. The technology captures sessions from more than 20 different AI coding tools and converts the useful ones into knowledge that agents can reuse later.
The issue is familiar to anyone working with AI coding assistants. A developer spends an hour debugging with Claude Code or Cursor, finally solves the problem, then starts a fresh session the next day only to watch the agent make the same mistakes all over again. Each interaction starts from scratch.
Beacon operates underneath these tools, recording every prompt, response, tool call, file edit, and approval in a unified format. After a session ends, developers can trigger an evaluation that analyzes the trace data and identifies which runs contain patterns worth preserving. That filtered knowledge gets stored locally and made available to agents through Model Context Protocol tools or as installable skill packages.
From logs to lessons
The technical architecture relies on OpenTelemetry standards to normalize agent activity across different platforms, according to company documentation. When evaluation runs, redacted session data flows to TypeSafe's Jev decision API or compatible alternatives, which return probability scores. Sessions need a task success score above 0.50 and a mean score exceeding 0.60 to become review candidates. Developers approve the most valuable runs, which then get written to a local memory.db file specific to each project.
The system works with Claude Code, Cursor, Codex, Cline, Copilot CLI, Devin, and roughly a dozen other coding harnesses. Everything stores locally by default in JSONL format. The GitHub repository has drawn attention from developers, though the company declined to share current installation numbers.
Justin D'Souza and Shukan Shah founded Asymptote after building detection systems and agentic security tools at Doppel. Both are based in New York. Tower Research Ventures led the funding round, joined by Lytical Ventures, 1984 Ventures, and South Park Commons. The company's current employee count was not disclosed.
"AI agents are becoming a new operating layer inside every company, but the infrastructure to understand and control them hasn't caught up," Asymptote said in its announcement. The founders positioned Beacon as open source deliberately, arguing that organizations should control their own agent data.
Tower Research Ventures framed Asymptote as building the "system of record for agentic work" and suggested that logs of agent activity will become critical enterprise assets.
Competing approaches
The memory problem has attracted multiple solutions in recent months. Mem0 introduced a token-efficient memory system in April 2024. Microsoft Research published PlugMem in March, focused on transforming interactions into reusable knowledge. ProjectMem arrived in June as a local-first option with event sourcing. Brain Memory launched in July with decay and salience mechanisms. Memoria, Memloom, Skillmem, and ai-memory offer similar engines.
Asymptote differentiates Beacon by emphasizing capture breadth. Rather than just storing memories, the platform records runtime telemetry across environments: local machines, CI pipelines, browser chats, cloud agents. The idea is to build a complete record of how agents actually work, not just what they remember.
The business model spans four tiers. Developers can self-host Beacon for free with a local dashboard. A free cloud version adds a unified timeline view. The Pro tier costs $20 monthly when billed annually and includes unlimited memory storage and installable skills. Enterprise customers get organization-wide policy controls and single sign-on, with pricing negotiated separately.
Whether enterprises will pay for agent memory infrastructure remains an open question. The technology addresses a real friction point, but the market for agentic tooling is still taking shape. Asymptote is betting that as companies deploy more AI agents, the ability to track and improve their performance will become table stakes. That assumes agents stick around long enough to make memory valuable, which seems likely but isn't guaranteed.
For now, the open-source strategy gives Asymptote distribution while it figures out which features warrant enterprise spending. The funding gives the team runway to find out.
