Memory that outlives the session
Decisions, errors, learnings and patterns in a shared MongoDB store — searchable by keyword or by meaning, across projects and across machines.
bee
The layer that keeps an AI coding agent oriented over a long run — persistent memory, behavioural rules, and per-file guidance delivered into the agent's context instead of its transcript.
Rules the agent actually receives
Behavioural rules injected at session start, per prompt, and on the file about to be written — not pasted into a prompt and forgotten three turns later.
Three runtimes, one tool
Claude Code, Codex and Hermes. Every integration was produced by running the real runtime and asking the model to quote what it received.
No secret in the binary
Provider keys, endpoints and connection strings live in the config store. The shipped tool contains zero credentials.
Why it exists
Section titled “Why it exists”An agent working for hours forgets. Not because its context is small, but because what it needs is not in it: the decision made yesterday, the convention that governs this one file, the reason the previous attempt was abandoned.
Pasting all of it into every prompt is the obvious fix and the wrong one — it costs tokens on every turn, and the fifty turns that do not need it pay the same as the one that does.
bee puts the right slice in at the right moment, and can tell you what that slice costs:
bee budget# session start (once)# ┌───────────┬───────┬───────┬─────────┐# │ component │ items │ chars │ ~tokens │# ├───────────┼───────┼───────┼─────────┤# │ memory │ 1 │ 108 │ 27 │# │ chrome │ │ 48 │ 12 │# │ total │ │ 156 │ 39 │# └───────────┴───────┴───────┴─────────┘