Memory

The memory agent keeps long-running work coherent across days or weeks.

What it does

It tracks durable decisions, assumptions, unresolved questions, and next-step priorities. It filters out transient chatter and preserves only information that should survive across sessions.

This allows research to continue without repeatedly rediscovering prior context.

Output

A structured memory log with decisions, active assumptions, open questions, and next-session bootstrap tasks.

Memory model

The memory agent organizes state into high-signal buckets:

• Decisions — what was decided and why
• Assumptions — what is believed but not fully validated
• Open questions — unresolved items blocking progress
• Next actions — concrete tasks for the next session

This format minimizes context loss while keeping memory concise.

What counts as durable memory

Good durable entries:

• model/provider defaults and why they were chosen
• accepted or rejected methodologies
• known data quality caveats
• repeatedly observed failure patterns

Do not store ephemeral notes that do not help future decisions.

When to use it

Use this agent when work spans multiple sessions, teammates, or iterations. It is especially useful for long investigations where priorities and assumptions evolve over time.

Limitations

Memory is a support layer, not a substitute for verification. Stored assumptions can become stale; contradiction and experiment passes should revalidate critical items.

Used by

Useful across all long-running workflows, especially /watch, /deepresearch, and ongoing replication projects.