Open-source developer tool · 2026
Continuum
A small, file-based continuity layer that lets AI coding agents pick up the same project state, decisions and preferences across tools and sessions.
- License
- MIT
- Platforms
- Windows, macOS, Linux
- Interface
- Plain Markdown and JSON
- Primary use
- Coding-agent handoff
01 The problem
Lost context
Coding agents keep their histories in separate tools. When a developer moves from Codex to Claude Code, returns the next day, or hits a usage limit, the next agent does not know the current task, recent decisions or the exact point where work stopped.
Re-pasting a long transcript is slow and noisy. Depending on one vendor's private memory also makes the project harder to move between tools.
02 Architecture
Portable ledger
Continuum keeps a compact .aicontext/ ledger inside the project. The live state, task list, decisions and journal stay readable by people and tools. Each supported agent receives the same protocol: read the ledger at session start, then update it before stopping.
STATE.mdholds the current focus, status, next steps and blockers.TASKS.mdseparates work in progress from backlog and completed work.DECISIONS.mdrecords why technical choices were made.JOURNAL.mdprovides a precise session handoff and left-off point.
Tool integration
Installers wire the protocol into Codex, Claude Code, Gemini, Cursor, Windsurf and other tools that read common agent-instruction files. Native hooks capture start, stop and pre-compaction events where a tool exposes them.
PowerShell and shell implementations mirror each other. Windows support is tested as a primary path, not treated as a later port.
03 Recovery and memory
Abrupt endings
If a session ends before it writes a handoff, Continuum can reconstruct the latest state from an agent transcript and git history. Drift checks compare the saved ledger with the repository so an agent does not present stale state as current fact.
Durable preferences
A separate global memory stores explicit, cross-project preferences. The next supported agent can apply the same development, design or communication preference without asking the developer to repeat it.
04 Outcome
What changed
Project context is no longer tied to one agent's private history. A developer can inspect the same state the agents read, edit it by hand, keep it local, or share it with a team through the repository.
The implementation stays intentionally small: files, adapters, hooks and installers, with no hosted account or required runtime service.