Shared memory for AI agents without shared confusion.

When should agents share memory?

Shared memory helps when agents work on the same subject over time: one gathers evidence, another reviews it, and a third continues the task later. Each agent should see the useful history without receiving every private note in the store.

A shared vector index is not enough. It can return similar text, but it cannot prove who wrote the record, who the record concerns, or whether the next agent is allowed to retrieve it.

• Use a stable subject for the person, system, or decision the memory concerns.
• Give every agent its own revocable credential and principal identity.
• Preserve the observer behind a claim instead of averaging different perspectives.

The boundaries that keep sharing safe

Titen stamps the actor from the credential and narrows records by organization, workspace, project, subject, agent, and run. Visibility is checked before ranking, so a record the caller cannot read contributes no score, conflict, or count.

Claims inherit the narrowest visibility of their supporting evidence. Consolidation therefore cannot turn a private observation into team memory.

Today, team and organization visibility behave the same inside one organization because workspace membership is stored but not yet joined into retrieval. That boundary is documented instead of hidden behind a broader promise.

Coordinate work without becoming a scheduler

Shared knowledge does not prevent duplicate work. Titen adds small coordination records while leaving agent selection, retries, and model loops to the caller.

• Checkpoints save expiring task state that another run can resume.
• Leases claim a resource for 10 seconds to 24 hours and return a conflict when another principal holds it.
• Handoffs name the target principal and can carry a checkpoint or compiled context.
• Leases are advisory. They do not fence database writes, so agents must agree to check them.

Keep disagreement as data

Two agents may observe different facts or hold different interpretations. A reliable shared memory system should expose that conflict rather than asking a model to invent consensus.

Titen keeps supporting, contradicting, and qualifying evidence attached to claims. Decisions replace earlier decisions through explicit supersession. Observer-scoped opinions remain separate.

The result is a memory layer a team can inspect: who knew something, which evidence supported it, when it was valid, and which context run used it.