Then, and today.
The apiary did not start from nothing. It is the continuation of Hivemind, the open-source shared-memory project from Activeloop, carried forward by Legion Code. This is the short story of how it got here and where it is going.
DownloadWhere it comes from
Hivemind set out to solve one stubborn problem: AI coding assistants are brilliant in the moment and forgetful the next. It gave a team one shared brain, so a fix one engineer found on Monday was available to every assistant on the team by Tuesday. It captured what assistants did, mined the repeating patterns into reusable skills, and shared those skills across the team, all on Activeloop Deeplake storage.
The apiary is what happened when two memory systems merged. Hivemind brought the broad product: capture across six and more assistants, the trace-to-skill mining, team skill sharing, a live map of your codebase, and the Deeplake substrate a team can share. Legion Code brought a stronger memory engine on top: a durable distill-to-recall pipeline, search by words and meaning together, memory kept at three levels of detail, a self-tidying loop, and source-backed recall that always shows where a memory came from.
Where the two overlapped on how memory should work, the Legion Code engine won. Then the family grew: honeycomb to remember, nectar to search code by meaning, hive to present it, doctor to keep the set alive, and queen to coordinate a fleet.
Milestones
- before the apiary
The forgetful assistant problem
Every AI coding assistant started each session cold. Close the window and the context was gone. The team could not see what any one person had learned. That gap is the reason both projects exist.
- apr 2026
Hivemind opens up
Activeloop publishes Hivemind as an open-source project: one shared brain for your agents, built on Deeplake. It captures sessions, codifies repeating patterns into reusable skills, and propagates them to every connected assistant. (github.com/activeloopai/hivemind, captured 2026-06.)
- spring 2026
Measured, not just claimed
On LoCoMo, a public long-context memory benchmark, Hivemind reaches the answer in fewer turns with less context than running with no shared memory: roughly 25 percent cheaper, about 1.7 times fewer tokens, and around 31 percent fewer turns. Prior work is already in scope at recall time instead of being re-derived each session. (github.com/activeloopai/hivemind, captured 2026-06.)
- the merger
honeycomb, the memory
Legion Code brings its memory engine: three-level memory, hybrid recall, the self-tidying loop, and source-backed provenance. The two systems merge, and honeycomb becomes the first product of the apiary, the one that remembers.
- the family grows
nectar, doctor, and hive
nectar adds meaning-based code search, so an agent finds the right file by what it does, not its name. doctor watches the local daemons and heals what breaks. hive gives the whole stack one always-on dashboard. Four roles on every machine.
- coming next
queen, the fleet orchestrator
queen is the cloud control plane for when the stack outgrows one machine. It sees and steers a fleet across machines and teammates, coordinating identity, presence, and ROI, without ever reading your memory content.
- now
The apiary, the continuation
The apiary is Hivemind carried forward: the same open-source roots and the same Deeplake substrate, with a sharper memory layer and five products on top. A Legion Code and Activeloop collaboration, where neither half stands alone.
What carries over, and what is new
The open-source roots carry over: capture across many assistants, skills that spread across a team, the live codebase map, and the Deeplake store you can keep in your own cloud. What is new is the memory layer underneath, the part that decides what is worth keeping, recalls it by words and meaning together, keeps it at three levels of detail, and shows you where every memory came from, plus the four other products that grew up around honeycomb. You get the breadth Hivemind proved out, with a stronger engine behind it.
Read about Deeplakebased on Hivemind benchmarks
One engineer's assistant figures out a tricky migration on Monday, and by Tuesday every assistant on the team can execute the pattern. That idea began in Hivemind and continues in the apiary.
Common questions
Where does the apiary come from?
It is the continuation of Hivemind, the open-source shared-memory project from Activeloop, carried forward by Legion Code with a stronger memory engine and a dashboard on top. Hivemind brought the breadth; Legion Code brought the memory layer underneath.
Is Hivemind open source?
Yes. Activeloop publishes Hivemind as an open-source project. The apiary builds on those open-source roots, and every product in the family is licensed AGPL-3.0-or-later.
What did Legion Code add?
The memory layer: a durable distill-to-recall pipeline, search by words and meaning together, memory kept at three levels of detail, a self-tidying loop, and source-backed recall that always shows where a memory came from. Then the family grew into five products.
Who is behind it?
A collaboration. Activeloop provides Deeplake and the open-source Hivemind roots. Legion Code adds the memory engine, the skill sharing, the self-tidying loop, and the daemon architecture that ties it together. Neither half stands alone.
See where the story leads.
Install the stack with one command and give your agents a memory that lasts.
Windows (PowerShell): irm https://get.theapiary.sh/install.ps1 | iex
Download