At the start of a session honeycomb hands your assistant a short briefing of what is already known about the project, so it starts informed instead of blank. It is small on purpose, just the headlines, so it never clutters the conversation.

how it works

Inside honeycomb

honeycomb runs one quiet loop underneath the assistants you already use: capture what happens, distill it into clean notes on Deeplake, recall the right ones next time, and compound by tidying itself as it grows.

Download Back to honeycomb

The loop, step by step

01
Capture

A small local daemon records the important moments of each agent turn: your prompt, the tool's actions, and the result. Capture is cheap and instant, and if it ever fails your assistant keeps working. Nothing is lost.
02
Distill

The daemon turns the raw transcript into a three-tier memory: a one-line key, a short summary, and the full original, so an agent can skim then zoom.
03
Recall

At the start of a session the assistant gets a short priming briefing. During work it can ask for more, and honeycomb matches both your words and the meaning behind them on Deeplake.
04
Compound

An optional self-tidying pass merges duplicates, drops junk, and keeps the current version of a fact, with full history, so memory gets sharper over time.

Skills that spread across your team.

When you or a teammate solve something reusable, honeycomb can turn it into a skill: a short, reusable lesson that shows up automatically in everyone's assistants at the start of their next session. Skills carry who wrote them, so credit and history stay clear, and promoting something from private to shared is always a deliberate, recorded action.

See security

recall "refresh token rotation"

auth/refresh-token.ts:42 ✓ verified

rotate the refresh token on every use; the old one is revoked server side.

auth/refresh-token.ts:42 personal 0.94

session:2f9c-deploy-staging

staging deploys run from the release branch, never from a feature branch.

session:2f9c-deploy-staging personal 0.89

db/migrations/notes.md:7 ✓ verified

add the column nullable first, backfill, then set not null in a later migration.

db/migrations/notes.md:7 personal 0.82

Under the hood

daemon: 127.0.0.1:3850
memory tiers: key, summary, raw
recall: word match plus meaning, hybrid on Deeplake
interfaces: CLI, MCP server, TypeScript SDK, harness hooks
storage: Activeloop Deeplake

Common questions

What is the three-tier memory?

Every memory is kept at three levels: a one-line key, a short summary, and the full original. An agent skims the key, opens the summary if relevant, and reads the raw detail only when it must.

What is priming?

At the start of a session honeycomb hands your assistant a short briefing of what is already known, so it starts informed instead of blank. It is small on purpose.

What is the self-tidying loop?

An optional pass that merges duplicates, drops junk, and keeps the current version of a fact while preserving history. It is why recall gets sharper as memory grows, and it is off by default because it uses a model.

See the loop for yourself.

Install honeycomb and watch a note written today get recalled tomorrow, in a different tool.

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

Download

Inside honeycomb

The loop, step by step

Capture

Distill

Recall

Compound

Skills that spread across your team.

Under the hood

Common questions

See the loop for yourself.