ContextLattice documentation#
ContextLattice is the local-first intelligence layer that gives AI agents durable continuity, explainable retrieval, portable context, and verified learning across harnesses.
This manual is generated from the Markdown in docs/wiki. Every page links back to its repository source, and every route is stable enough to bookmark, share, or open directly in an agent.
Choose your path#
- Install and verify the runtime if this is your first machine.
- Understand the operating model before tuning retrieval or storage.
- Use the primary CLI for the normal context lifecycle.
- Connect an agent or application after the core runtime is healthy.
- Operate and recover the stack with bounded checks.
- Diagnose a failure from the exact failing boundary.
- Review release posture before upgrading or publishing.
The shortest useful loop#
zsh
contextlattice context "what must I know before this task?" --project my-project --pretty
contextlattice remember "verified checkpoint" --project my-project --pretty
contextlattice finish "work completed and tested" --project my-project --success --prettyUse contextlattice resume --project my-project --pretty when a later session needs the bounded objective, evidence, risks, and next action. Use contextlattice correct when retrieved context is wrong, stale, useful, or superseded.
Interface order#
- CLI is the primary human and agent interface.
- Dashboard is the visibility, proof, account, and operations companion.
- HTTP and MCP are integration surfaces for harnesses and applications.
The command line keeps the normal workflow explicit. Programmatic clients should use the same scoped lifecycle and avoid replaying full transcripts as memory.
What good context looks like#
- Scoped to a project, topic, task, or session.
- Small enough to inspect before acting.
- Grounded in source paths, timestamps, identities, or command evidence.
- Honest about missing or degraded sources.
- Correctable when evidence changes.
- Finished with a verified outcome so retrieval quality can improve.
ContextLattice should make an agent more continuous, not more credulous. Retrieved records are evidence to inspect, never authority to expand a task or override current instructions.