Project memory

What a memory is

A memory is a short note about a project that's worth knowing next time, such as a decision and its reason, an approach that failed, or a command that works. OpenOrc adds the best-ranked ones to the instructions of each new run in the project, so agents start from what earlier runs learned.

Each memory has a type, a title, a body, and a confidence from 0.1 to 1, and it can name the files it concerns. The types are Decision, Lesson, Command, Env quirk, Convention, Preference, Spec, and Ownership. Memories belong to a project, and every agent working there can read them. Orclings keep a private memory of their own; see Orclings.

Turn memory on

Memory is off in a new profile. One switch turns it on for all your projects.

1. Click Settings at the bottom of the sidebar, then Memory & models.
2. Turn on the switch next to OpenOrc memory is off.
3. Read the line under the heading. It names the agent and model that will summarize your finished runs.

Turning memory off stops OpenOrc from saving or recalling memories. The memories you have stay, and you can still edit and delete them.

How memories are made

Memories come from two places:

• After a run, a small model reads a digest of it and saves up to six memories, plus a summary of what was asked, what was done, and what was left open. These memories are marked Learned after a run.
• An agent can save one while it works. These are marked Saved by an agent, and the transcript shows a Remembered step.

A run is over when the agent's process exits. In a conversation, that happens once the agent has been idle for the time in Close idle processes after (Settings → General, 10 minutes by default), so a conversation is usually summarized once per working session.

When a new memory has the same topic as an existing one, it updates that memory and raises its confidence by 0.1.

Review and edit memories

1. Press ⌘K and choose Memory.
2. Choose a project in the menu next to Memory at the top of the page.
3. Type in Search memories…, or filter by type and by who saved them.

Each memory shows who saved it and when it was created, updated, and last confirmed. Its three-dot menu has these actions:

In a conversation, the side panel's Memory tab lists the project's memories once there are any. A task's Memory tab shows Run history and Memory from this task instead. Workspace conversations, Slack ones included, share one memory that the project menu doesn't list; open the Memory tab in a Workspace conversation to see it.

How later runs use memory

When memory is on and the project has memories, OpenOrc adds a project memory section to the brief each run gets. It holds up to 6 decisions, 8 lessons, 6 commands, 4 env quirks, 5 conventions, and 5 preferences, ranked by confidence and by how recently each was confirmed. Spec and Ownership memories stay out. The section also lists tasks in progress or in review, and items that recent runs left open. Each line shows its age and who saved it, and the section stops at 5,000 characters.

OpenOrc builds the brief when an agent's process starts, so a new memory reaches a conversation the next time its agent starts. During a run, agents can also search memory by words and by meaning with memory_search, save a memory with memory_record, and mark one helpful, wrong, or stale with memory_feedback. A run on a task can call task_context for the task's spec, its earlier runs, and the most relevant memories.

Saving and rating follow the conversation's mode: Plan blocks them, Review everything asks you first, and Accept edits and Autonomous allow them. See OpenOrc's tools. While memory is off, agents get only task_context.

Choose who summarizes runs

In Settings → Memory & models, the Learn from completed runs section decides who writes the summaries. With memory on, choose a Provider:

Except with an API key, the agent that summarizes must be signed in. The line under the section says who will summarize runs, or why no one can. Text generation, on the same page, picks the model that names new threads and works separately from memory.

What leaves your computer

• Memories, run summaries, and the search index stay in OpenOrc's data folder. Searches run on your computer.
• With a cloud brain, the memory of each project you switch on for it is shared with the brain on your cloud account, and merged both ways while both are awake.
• Searching by meaning needs a small embedding model. The first time memory needs it, OpenOrc downloads it (about 91 MB, from Hugging Face) and checks each file against a fixed checksum. Nothing is downloaded while memory is off.
• Each summary sends a digest of the run to the summarizing provider under your account: your messages, changed files, recent tool calls, and the agent's last messages, up to 12,000 characters. It comes from OpenOrc's stored copy of the run, which has been through a best-effort filter for well-known secrets such as API keys.
• With Automatic · per agent, a run's digest goes only to the provider that ran it. A provider you choose receives every run's digest.

Next steps