honeycomb

AI coding agents forget. They forget across sessions, and they forget across tools. A decision you reached in Claude Code at midnight is invisible to Cursor the next morning. Honeycomb fixes that. A local daemon captures what happens on every turn, distills it, and serves it back to any harness that asks. Learn something once; recall it everywhere, on any machine, in any tool, for anyone on your team.

New here? One command and you're on a dashboard. Jump to Install. · Want the docs? Everything lives at theapiary.sh.

🛹 For AI augmented devs

Stop re-explaining your project to a fresh agent every morning. Honeycomb remembers your decisions, your conventions, and the fixes that worked, then primes your next session with them automatically. One install command, a friendly dashboard, no SQL, no config gauntlet.

🏢 For enterprise teams

One shared brain across every developer, device, and coding tool. A skill discovered by one engineer propagates to the whole team on their next session. Tenancy is enforced at the storage layer; credentials live behind a single loopback boundary; everything is versioned and auditable.

✨ What makes Honeycomb different

A vector database can store text and hand it back by similarity. Honeycomb does that, and then keeps going. On top of Activeloop Deep Lake, Legion Code builds the memory system that turns raw recall into a brain your agents actually trust:

• 🧠 Three-tier memory. Every memory exists at three resolutions at once (one-line key → summary → full raw session). Agents skim the keys, then zoom into detail only when they need it. (Legion Code)
• 🎯 Session priming. At session start a tiny, bounded index (~300-800 tokens) of your most relevant keys is pushed once; the agent pulls deeper on demand. No per-turn injection, no "lost in the middle." (Legion Code)
• 🍯 Skillify & propagation. The daemon mines reusable skills out of real sessions, gates them for quality, and auto-pulls the team's latest skills into every agent at session start. Author a skill once; everyone gets it. (Legion Code)
• 🌼 The pollinating loop. A periodic maintenance pass reasons over accumulated memory and the entity graph to merge duplicates, prune junk, and supersede stale facts, so memory gets sharper over time, not noisier. (Legion Code)
• 🕸️ Knowledge graph. An entity-centric, versioned, provenance-tracked index over your memories. Newer facts supersede stale ones; every claim traces back to the session that produced it. (Legion Code)
• 🔀 Hybrid recall. Lexical (BM25) and semantic (768-dim vectors) search fused by Reciprocal Rank Fusion, with a measured recall@5 ≈ 0.72-0.78. (built on Deep Lake)
• 🗺️ Codebase graph. A multi-language AST graph (TypeScript, JS, Python, Go, Rust, Java, Ruby, C/C++) of files, functions, and their call/import/extends edges, queryable for impact and neighborhood. (Legion Code)

🚀 Install (one command)

No Node? No npm? No problem. The installer detects and sets up everything, then opens a dashboard in your browser. The terminal is just a progress log; the product is the first thing you touch.

# macOS / Linux
curl -fsSL https://get.theapiary.sh | sh

# Windows (PowerShell)
irm https://get.theapiary.sh/install.ps1 | iex

That single line installs a current Node/npm if missing, installs @legioncodeinc/honeycomb globally, brings up the daemon on 127.0.0.1:3850, opens the dashboard, and sets up HiveDoctor, a tiny watchdog that keeps it all healthy (opt out with --no-hivedoctor). Then:

1. The dashboard loads in a pre-auth setup state. No token ever touches your shell.
2. Click "First time setup." Honeycomb runs the Deep Lake device-flow login for you, shows the code right on the page, and opens the verification tab.
3. Done. The same running daemon lights up its Deep Lake-backed surfaces, and capture and recall go live.

Already running Hivemind? The dashboard detects it, explains that running both is unsupported, and "Proceed with Honeycomb" migrates you cleanly. Prefer to inspect before you pipe? The script and a published SHA256SUMS are served from get.theapiary.sh.

git clone https://github.com/legioncodeinc/honeycomb.git
cd honeycomb
npm install
npm run build          # tsc + esbuild → bundle/cli.js, daemon, harness, MCP, embed bundles

node bundle/cli.js setup     # detect your assistants, wire hooks, start the daemon
node bundle/cli.js status    # check the daemon and your environment

setup wires every coding assistant it detects and starts the loopback daemon; any storage command auto-starts the daemon if it is down. You'll need Activeloop Deep Lake credentials; the device flow above writes them to the shared ~/.deeplake/credentials.json.

🐝 First memory, shared across tools

# Capture a decision once…
honeycomb remember "we deploy from the prd-022 branch, never from main"

# …recall it anywhere: same daemon, same Deep Lake, any harness
honeycomb recall "how do we deploy"

Write it from Claude Code; recall it from Cursor tomorrow on a different laptop. That's the whole point.

🏗️ How it works

Honeycomb is a long-lived local daemon plus thin clients. The daemon is the only process that talks to storage. Every harness, the CLI, the MCP server, and the SDK reach it over loopback HTTP. One shared memory behind one boundary; your Deep Lake credentials in exactly one place.

     Claude Code    Cursor    Codex    Hermes    pi    OpenClaw
          │           │         │        │       │        │
          └───────────┴────┬────┴────────┴───────┴────────┘
                 hooks · CLI · MCP · SDK   (thin clients)
                           │
                  loopback HTTP · 127.0.0.1:3850
                           │
            ┌──────────────────────────────┐
            │       honeycomb daemon        │   sole storage client
            │  capture · recall · skillify  │   owns your credentials
            │  pollinate · session priming  │
            └───────────────┬──────────────┘
                            │
            ┌──────────────────────────────┐
            │     Activeloop Deep Lake      │   versioned · columnar + vector
            │   Tier 1 · Tier 2 · Tier 3    │   BM25 + semantic hybrid
            └──────────────────────────────┘

• Capture on every turn. Per-harness hooks stream each turn to the daemon, which distills and persists it: always-on, cheap, and soft-failing so a capture error never breaks your agent's turn.
• Recall through the daemon. Any harness asks for relevant memories; the daemon runs the query and returns results already scoped to your org and workspace. The client never sees a storage handle or a line of SQL.
• Shared by construction. Every client reaches the same daemon and the same dataset, so a memory written from one harness is recallable from all of them.

🧠 The three-tier memory system

This is the heart of what Legion Code adds on top of Deep Lake. The same memory lives at three levels of detail at once, and the agent chooses how far to zoom: