Shared memory for AI agents: project scope, ownership and conflicts

The lost update is easy to miss

Two agents open the same project note. One corrects a deadline. The other adds a source and saves the older copy it read. Both report success, yet the deadline correction is gone. They shared storage without coordinating their edits.

Before connecting a second writer, define ownership and how changes become current. If agents only need to answer questions, a read-only shared view may be enough. Add concurrent writing only when there is a reason to support it.

Decide who can edit what

These patterns need support from your storage and agent tools. They do not demonstrate distributed locking or automatic conflict resolution in Gigabrain.

Carry the version through the handoff

Agent A owns: project-summary.md
Agent B owns: research-notes.md
Starting revision: Record the actual version or file hash
Shared sources: Read-only approved source folder
Cross-file changes: Propose; do not overwrite
Before saving: Check whether the owned file changed
After saving: Record change, evidence and new revision
Conflict: Stop the write and compare both versions

The version check must occur through a supported mechanism. A check performed long before a write can still race with another change. For consequential concurrent writes, use storage that enforces the coordination you require rather than relying on agents to remember a convention.

Understand what Git worktrees solve

Git worktrees allow multiple working trees associated with a repository. They can help agents keep edits in separate working directories and branches. They do not decide which factual change is correct, grant privacy between agents or make every concurrent operation safe.

Read the merged text and check what it says, even if Git reports no conflict. A text merge can succeed while combining an old deadline with a new decision. Check the merged facts against their sources and keep a record of why you chose that resolution.

Test a deliberate collision

Use a disposable note with two independent fields. Have two controlled sessions propose changes from the same starting revision. Confirm that the second change is reviewed against the first and that neither disappears. Then repeat with conflicting values for one field.

Success means the conflict is surfaced and resolved explicitly, not that one agent happens to save last. Keep the proposed versions and the resolution record. Never run the experiment on the only copy of active project knowledge.

Start with a shared view you can inspect

Gigabrain is intended for durable, connected project knowledge. For a multi-agent setup, verify each integration and define the write boundary before sharing the workspace. A licence quantity does not itself provide technical tenant isolation or conflict management.

Use the handoff and ownership template, inspect current evidence and review licence options for the people who need the system.

Resolve two proposals from the same starting note

In this fictional exercise, the planning agent and review agent both read revision 7 of Alder's inspection note.

This is an ownership exercise, not a measured concurrent-agent run. A clean text merge alone would not show that the date is current.

Use the resolution record to preserve both proposals, the starting version and the accepted facts.

Test the scope separately from the write convention

The component checks test that a scoped graph read cannot follow another project's sources. They also test approval binding and changed proposals.

A project-scoped connection does not prevent an agent with separate filesystem access from reading files its host permits. Gigabrain's licence count does not establish hosted tenant isolation.