ContextLattice is a local-first intelligence layer for AI agents, providing agents with persistent continuity, explainable retrieval, portable context, and verified learning capabilities across frameworks. The problem it addresses is this: models can reason, and agent frameworks can act, but when chats, models, tools, accounts, or computers change, neither can reliably preserve task goals. ContextLattice provides a persistent, inspectable context layer for this kind of work, reconstructing the current goal, selecting important evidence, carrying it safely, and recording what actually worked, while avoiding turning every prompt into a full conversation transcript dump and without forcing the use of cloud storage.
In terms of main capabilities, it provides a persistent memory lifecycle, Context Packs, sessions, graph and claim interfaces, Skills Index discovery, Agent Packets, public Passport and Mesh contracts, and local attestation tooling. Retrieval merges available sources, ranks limited evidence, and reports missing or degraded coverage rather than hiding these issues. Writes are verified and durably saved before distribution. It supports staged retrieval, fused data backends, and long-horizon context continuity.
In terms of usage, the CLI is the primary interface, the dashboard makes behavior and evidence visible, and HTTP and MCP are companion integration surfaces for applications and frameworks. Installation requirements are macOS, Linux, or Windows via WSL2; a container runtime compatible with Compose v2; and gmake, jq, rg, python3, and curl. The tested macOS baseline uses OrbStack through an explicit Docker context.gmake quickstart is the prescribed technical installation path, and the installer is the guided alternative. This command prepares environment wiring, asks about runtime configuration, starts the selected local stack, and verifies initial readiness. A healthy container is only the first step,state Command verifies the canonical storage manifest owned by the gateway; subsequent commands execute the actual context path and report source coverage, degradation, evidence, and follow-up actions.
To connect an agent, run the integration command in the repository where ContextLattice should be used. This command writes a managed boundary block while preserving existing instruction text, but it does not install Codex, Claude Code, OpenCode, Hermes, OMP, Mercury, Pi, Droid, or other third-party agent frameworks. External provider discovery is network-free, and provider execution remains explicitly authorized by the caller.
Applicable scenarios include developers and teams who want to maintain task continuity across different agent frameworks, models, or sessions, as well as scenarios that require explainable retrieval, evidence provenance, and portable context without being forced to depend on cloud storage. It is suitable for local-first, privacy-focused, and controllability-oriented workflows.
Regarding limitations and boundaries, ContextLattice does not replace agent frameworks, does not select targets from retrieved text, and does not execute imported context. Local tools remain the execution surface, and memory and remote content remain only evidence. Continuous Cognition is advisory only: each invocation issues one bounded request, returns opaque evidence references, and never dispatches a runner or performs external changes. Context preparation declarations are kept only in owner-only accessible files and cross the completion/failure boundary only in protected headers. Under the default local control path, writes are distributed only after validation and persistence. Secret-like content is deterministically filtered at the write entry point: redacted by default, blocked when configured, and allowed only when explicitly permitted by the operator. Retrieved memories have provenance and trust isolation. Portable context and continuation envelopes are signed and encrypted. Graph repair, provenance backfill, and isolation workflows default to dry-run. No retrieved instructions or imported continuation content are automatically executed. The public local channel has no account and can be used independently; paid artifacts add governed collaboration, protected activation, workspace operations, advanced analytics, and managed distribution, but do not turn local memory into a mandatory cloud dependency. The current release baseline is v5.0.2, and the public channel is licensed under the Apache License 2.0. Specific pricing and plan details are not yet provided.
