Memory

Memory is what turns a chatbot into a teammate. Every agent in Busy Bee remembers what's happened, what's been decided, and what's likely to come up again.

If agents started every task with a blank slate, you'd spend half your day re-establishing context. Busy Bee's memory layer keeps relevant history alive across runs, without making you manage it by hand.

Three layers of memory

An agent draws on three memory layers, in order of scope:

Conversation memory: the running thread of the current task. Always available, never lost mid-task.
Project memory: durable across tasks within a project. "What's been shipped, what's been decided, what context matters." Agents working in the same project see the same project memory.
Agent memory: durable across all tasks the agent has touched, across all projects. The agent's accumulated experience.

The three compose: when a task starts, the agent has the conversation (full), the project memory (summarized), and its own long-term memory (queried by relevance).

What gets remembered

Memory entries fall into a few buckets:

Decisions - "We decided to use Postgres over MySQL." The agent stores the decision and the reasoning, so the next time someone asks "why are we on Postgres?" the agent can answer.
Constraints - "Never include competitor names in marketing copy." Caught once; never violated again.
Preferences - "BK prefers bullet lists over prose for summaries."
Facts - "Our production region is us-east-1." Stored once; referenced everywhere.
Open questions - "Pending answer: which billing provider are we using?" Surfaces back to the agent on the next relevant task.

Agents write to memory autonomously, but they tell you what they're storing. You can review and edit memory entries from the agent's settings page.

Memory scope and access

On an agent's settings page, you control:

Which projects the agent has memory access to.
Whether the agent can write to long-term memory, or is read-only.
Whether memory is shared across agents in the same team (default: shared within a Hive, isolated across Hives).

The default scope is sane for most workflows. Tighten it for security-sensitive contexts.

📸 Screenshot: an agent's memory inspector showing recent entries with edit and forget controls.

Forgetting

Two ways to remove a memory entry:

Edit - change the content of an entry the agent has already stored. Useful when a decision changes.
Forget - remove the entry entirely. The agent will no longer surface it, and won't recreate it unless re-encountered.

There's no "wipe everything" button by design, too risky. Forget specific entries when you need to.

Memory and privacy

Memory entries live in your workspace database, scoped to your organization. They're not shared with other Busy Bee customers, not used to train models, and can be exported or deleted on request. Sensitive entries (credentials, PII) shouldn't end up in memory, and the agent is trained to avoid storing them, but audit your memory entries periodically all the same.

What's next

Memory is what makes individual agents better over time. Teams are how a group of memorable agents becomes a coordinated unit.