We do the model homework.
COLLASTRA breaks the goal into tasks and matches each one to an agent or model’s capabilities, with your constraints in view.
HOW COLLASTRA WORKS
One goal.
The whole team behind it.
COLLASTRA assigns the work, carries the context, and keeps the lessons you review.
Plan → Coordinate → Review → Remember
ONE GOAL“Fix the invitation retry.”Illustrative workflow
✦ The team’s task plan
COLLASTRAOne goalRunning
Fix the invitation retry.
Claude CodeInvestigate
Find why a retry invalidates a valid invitation.
CodexImplement
Preserve the token and add a retry state.
Claude CodeReview
Check expiry, recovery and duplicate submissions.
Select a stage to follow the work.Shared memory / OctaMem
No handoff from scratch.
The goal, relevant files and prior decisions move with the work. The next agent can pick up where the last one stopped.
Experience that compounds.
Review the result. Keep the useful lesson in OctaMem. Future runs can reuse what worked, even with a different agent.
A little more clarity.
Does shared memory train the models?
No. Reviewed lessons are context supplied to agents on later tasks. The underlying models are not retrained.
Are efficiency gains measured?
Not yet. Less repeated work and unnecessary token use are design goals. We will publish measurements when we have them.
Which connections are supported?
Claude Code, Codex, Hermes Agent and Google Antigravity are part of the beta team. Models are available through OpenRouter. Cursor, OpenCode and further tool integrations are coming soon. The website uses illustrative examples.