Local-first, governed memory for Claude Code, Codex, and Cursor — shared across every session and fully yours, so your agents stop re-breaking what you already fixed. Debug it. Prove it. Purge it.
Every session starts cold. The bug you already squashed, the convention you agreed on, the fix you shipped yesterday — gone by tomorrow. A CLAUDE.md rots and drifts out of date; a cloud memory tool hides what it knows and you can't audit it. So the same mistakes keep coming back.
isanna Hive is a local-first memory substrate for AI coding agents — persistent, searchable, and fully governable. A free, open-source core (Apache-2.0), plus a premium Console for teams who need to trust what their agents know.
Every search runs two retrieval arms in parallel and fuses them — semantic (meaning) and keyword (exact term) — so recall doesn't miss what pure vector similarity would.
Honest about what's on. Vector + keyword fusion (RRF), the relevance/recency/importance blend, and recall telemetry are default-on for every search — no flag. Opt-in (real, gated): point-in-time recall (HIVEMIND_PIT_RECALL_ENABLED). Roadmap — not yet implemented: structured-table recall, graph-augmented recall, a learned reranker, and reflective retry.
Published agent research keeps finding the same thing: memory of past work measurably improves task success and cuts repeated mistakes.
Reflexion · 91% vs 80% Agent Workflow Memory · +51% Voyager · 15× faster MemGPT · memory OS CoALA · agent memorySeven views, grouped into three simple moves — everything your agents remember, under your control.
Under the hood, recall blends vector, keyword (RRF), structured, and graph signals into a hybrid rank based on relevance + recency + importance. Full implementation evidence is shared directly with reviewers and beta buyers.
Yes. Hive Core is free and open source under Apache-2.0 (including an explicit patent grant) — self-host the full memory substrate forever, no strings and no telemetry.
On a box you own. Hive is local-first — your agents' memory text isn't routed to an outside service.
Anything that speaks MCP — Claude Code, Codex, Cursor, and more — reading and writing through one governed seam.
No. The Console brings up Postgres + pgvector in Docker for you — one image, identical on macOS, Windows, and Linux.
The governance layer — debug, prove, and purge what your agents know — as a read surface over the same local store.
Start with the free core, or apply for the private beta.