Memory and decisions

Agentbits keeps a long-term memory so the lead and agents don’t start from zero on every run. Memory records are short pieces of context: what a task produced, what the lead asked to fix, what you said in chat, which decisions were accepted. Before an agent works, the most relevant records are added to its prompt.

Memory is stored in the local database on your computer, with full-text search (SQLite FTS5). Nothing is sent anywhere except inside the prompts of the provider an agent uses.

Project memory and organization memory

Every record has a scope:

Scope	Where it is used	Typical content
Project memory	Only in that project	Task summaries, review notes, chat context, accepted decisions
Organization memory	In every project	Your general preferences and shared rules, for example “Keep UI copy short; code comments in English.”

When an agent works in a project, it sees that project’s memory plus the organization memory. It never sees another project’s memory, so context from one client or codebase does not leak into another.

A record can also be shared by all agents or belong to one agent only. Review feedback, for example, is stored for the agent whose work was sent back.

What gets written to memory

Agentbits writes these records automatically, all to the project’s memory:

• Task summaries: when an agent finishes a task, a summary of its report.
• Review feedback: when the lead sends a task back, its note.
• Your chat messages: messages of 20 characters or more that you send in the project.
• Accepted decisions: with the highest importance, so they are always considered.

You add records yourself in Settings → Memory:

• Pick the scope: Organization-wide or This project.
• Write the record, for example a rule the whole team should follow.
• Choose who it is for: Shared (all agents) or one agent only.
• Save it with Add to organization memory, or with Add to project name memory for the project scope.

The same screen lets you search memory and delete records.

How memory enters the prompt

Each time an agent starts a task, Agentbits searches memory using the task’s title and description. Chat replies draw on the project’s memory too. The matches are ranked by:

• Relevance to the text counts most (60%).
• Importance counts next (30%). Accepted decisions have the highest importance, review notes and task summaries less.
• Recency counts least (10%). Newer records rank higher, and the effect fades over a couple of weeks.

Records with very high importance, such as accepted decisions, are included even when no words match. The top records, up to eight, go into the agent’s system prompt in a section called “Relevant memory”, which tells the agent to treat them as ground truth unless something contradicts them.

You can see exactly what an agent received: in Terminal → Agent terminals, open the Full prompt sent entry for a run and look at the System prompt. See Terminals.

Note: Memory uses text search, not semantic (embedding) search, which is planned. A record that uses different words from the task may not be found. Keep important rules short and use the words agents will see in tasks.

Decisions

Decisions are architecture-decision-style records (ADR) with a Context, a Decision and a Rationale. Each project has its own Decision Log in the Decisions section (Ctrl+7).

A decision has one of four statuses: Proposed, Accepted, Rejected or Superseded.

How decisions are made

1. An agent or the lead proposes it. Agents are asked to report important decisions in their output, and the lead can propose them in chat. These arrive in the Decision Log as Proposed, a counter appears in the top bar, and a message in the chat points to them.
2. You decide. Open the decision, edit it if needed, and click Accept, reject it, or later Supersede it with a newer decision. You can also delete it.
3. Accepted decisions go into memory. An accepted decision is written to the project’s memory with the highest importance, so it reaches every agent working in that project from then on.

You can also write decisions yourself: click New, fill in the fields and choose Propose or Accept directly.

Use the filter (All, Proposed, Accepted, Rejected) to find decisions quickly.

Clearing memory

• Settings → Data management cleans up a single project: its project memory, chat, run records, reports, decisions, tasks or files, in any combination. Tick Keep decision memory to keep accepted decisions while clearing the rest.
• The same screen clears the Organization-wide memory, with a Keep decisions option.
• In the project settings, Clear memory and history → Manage data opens this screen with the project already selected.
• Deleting a project deletes its project memory. Deleting an agent deletes its personal memory.

See Settings for backups and resets.

Tip: Put rules that apply everywhere (“answer in English”, “never touch the production database”) in organization memory, and project-specific facts in project memory or as an accepted decision.