Wyrm is a local-first memory and intelligence layer designed to provide AI agents with persistent, searchable memory across sessions. It enables tools like Claude, Cursor, and Copilot to retain context, learn from failures, and prevent the repetition of known mistakes. Operating entirely on the user's machine, Wyrm ensures that no data leaves the local environment by default, offering a private and secure solution for AI memory management.
Key features of Wyrm include: Persistent Memory (projects, sessions, quests, skills, and ground truths, all kept in one local SQLite file); Hybrid Recall (full-text search fused with semantic vectors, local by default via a bundled model, with optional NVIDIA NIM embeddings and reranking); a Benchmarked Firewall (negative learning that blocks a repeated mistake, measured at 100% recall and 100% precision, sub-millisecond, reproducible from the public repo); Fleet Memory and Fleet Negative Learning (every write attributed to the agent and run that produced it, so a swarm of agents shares one accountable, mistake-blocking memory bus); a Knowledge Graph (entity and relationship traversal, pathfinding, and merging); Portable Skills (SKILL.md content synced across every machine); a Render Target (wyrm render compiles memory into CLAUDE.md, AGENTS.md, and other agent context files, deterministically and at zero added token cost); Provable Memory (every write returns a receipt: stored, queued, merged, or dropped, and why); and Provenance Trust Lanes (untrusted content is quarantined from context briefs so a planted instruction can never be read back as a command).
Wyrm installs via npm, works fully offline with no account or login, and is free to use with no usage limits on the local tier. Paid plans add cloud sync across devices, encrypted snapshots, shared team memory, and support. Built by Ghost Protocol (Pvt) Ltd in Colombo, Sri Lanka, and a member of the NVIDIA Inception program.
Who Is the Company Behind Wyrm?