Concept
Persistent Project Context for AI Agents
Keep project decisions, constraints, and handoffs available across Claude Code, Cursor, Codex, and other AI tools.
Article packet
Concepts
Developers using multiple local AI tools on multi-session projects
5 min read
01
Persistent context is more than a saved prompt.
02
It includes decisions, constraints, gotchas, project status, and source-backed pages.
03
Origin keeps that context local and retrievable through MCP.
01
Quick answer
Persistent project context is the durable working state an AI agent needs to continue a project later: what changed, what was decided, what to avoid, and where the source of truth lives.
Origin stores that local work context as memories, session handoffs, project status, and distilled pages. It is not full transcript archival; MCP clients retrieve the relevant pieces instead of forcing the user to paste a project briefing every time.
02
When this problem appears
Projects rarely fail because the agent forgot generic programming knowledge. They fail because it forgot local decisions: which tradeoff won, which migration was risky, which command proved the fix, or which feature is out of scope.
03
Capture project context
Focus memory on a project context packet a future agent cannot infer cheaply.
- Start with /brief in Claude Code or Origin context from an MCP client.
- Capture decisions and why they were chosen.
- Capture gotchas that would waste time if rediscovered.
- Capture source-of-truth files, current status, stale facts, and constraints when they are not obvious from the repo.
- Use /recall for specific prior decisions instead of asking the model to guess.
- Capture project status at session end with /handoff or a handoff-style MCP capture.
- Distill repeated clusters into pages when a topic keeps returning.
Claude Code daily loop
/brief
/capture <one durable project fact and why it matters>
/recall <specific prior decision or gotcha>
/handoff04
What to check next
Do not store facts that the repo can read directly unless the important part is the reasoning behind them.
Try the local memory loop
Install Origin, connect your AI client, and verify that capture, recall, and handoff work on your machine.
FAQ