Until yesterday, agent memory was a black box. Our agents were reading and writing memories, but we had no way to see what they remembered, correct mistakes, or seed useful context. That changed with Cycle 8.
What We Shipped
The Memory Observatory is a web interface and API for working with agent memory. You can now:
• Browse all memory entries by scope (global, task-specific, or custom)
• Edit existing memories to fix errors or update facts
• Inject new memories to give agents context they need
• Delete stale or incorrect entries
Under the hood, we built REST API endpoints (/api/memory/browse, /api/memory/[id]) with proper authentication middleware and comprehensive test coverage. The tests validate CRUD operations, auth flows, and edge cases.
Why This Matters
Agent memory is how our system learns and retains context across tasks. Before this, if an agent learned something wrong — say, an incorrect file path or outdated API structure — we couldn't easily fix it. The agent would keep using that bad information.
Now we can observe what agents remember and intervene when needed. We can also seed memories proactively: onboarding docs, architecture decisions, team conventions. Instead of waiting for agents to discover these through trial and error, we can tell them upfront.
What's Rough
This is a v1. The UI is functional but minimal. There's no search or filtering yet, so browsing large memory scopes is tedious. We don't have version history, so if you edit a memory, the old value is gone. And there's no bulk import — you're injecting memories one at a time.
We also haven't built analytics or insights. You can see what's stored, but not how often it's accessed, which entries are stale, or what memories correlate with successful task completions.
What's Next
Near term: search and filtering. We need to be able to find memories by keyword, scope pattern, or confidence score. We're also adding CSV import so you can seed memories in bulk.
Medium term: memory analytics. Which memories get accessed most? Which have low confidence but high access counts (indicating uncertainty)? Which scopes are growing fastest? This will help us understand how agents learn and where they struggle.
Longer term: collaborative memory editing and version control. Right now, memory is a single-player tool. We want multiple people to be able to curate agent knowledge together, with change tracking and rollback.
For now, the observatory is live. If you're working with agents, take a look at what they remember. You might be surprised.