What it does

Hivemind is a shared memory layer for AI coding agents. It hooks into Claude Code, OpenClaw, Codex, Cursor, Hermes, and pi to capture every prompt, tool call, and response as structured traces. A background worker mines these traces for repeated patterns, codifies them into SKILL.md files, and injects them back into every connected agent’s context. The pitch: your senior engineer’s agent figures out a migration on Monday; your junior engineer’s agent executes the same pattern on Tuesday without re-deriving it.

The interesting bit

The “virtual filesystem” trick. Hivemind intercepts file operations on ~/.deeplake/memory/ through a SQL-backed virtual filesystem, letting agents read and write shared memory as if it were local files. The retrieval stack also falls back from semantic search to BM25 when embeddings are unavailable — a pragmatic admission that vector search isn’t always on.

Key highlights

• Benchmarked on LoCoMo: 25% cheaper, 1.7× fewer tokens, 31% fewer turns versus no-memory baseline (Claude Haiku, hybrid lexical + semantic retrieval)
• BYOC storage: GCS, Azure, S3, or on-prem buckets — data doesn’t sit in Hivemind’s infrastructure
• Real-time propagation across sessions, agents, teammates, and machines
• Auto-generated wiki pages summarizing sessions at session end
• Y Combinator-backed; built by the Deeplake team

Caveats

• The LoCoMo benchmark numbers are for Claude Haiku specifically; performance on other models is unstated
• Several integrations (Cursor 1.7+, Hermes, pi) require hooking into relatively new or unstable extension APIs
• The README’s “How it works” section is truncated, so the full mechanics of skill codification and conflict resolution are unclear

Verdict

Worth a look if you’re running multiple AI agents across a team and tired of watching them rediscover the same codebase quirks. Skip it if you’re a solo developer on a single agent — the network effects are the whole point.