ContextLattice

What sheawinkler/ContextLattice MCP server does

sheawinkler/ContextLattice MCP server provides durable memory and context continuity for agents. It stores and reconstructs an active objective together with decisions, repository state, risks, proof, and the next action. Retrieval is bounded rather than transcript-based: the system selects high-signal evidence into a compact context packet and reports provenance, source coverage, omissions, contradictions, and degradation.

The project is local-first and private by default. Raw memory remains local, while signed and bounded continuation data can be moved between agents or machines. The repository describes a Go/Rust runtime, fused data backends, staged retrieval, and long-horizon continuity. A dashboard exposes behavior and proof, but the CLI is the main operating surface.

How it works

The documented workflow has five stages:

1. Reopen reconstructs the current mission from durable checkpoints and present state.
2. Select retrieves relevant evidence into a compact, provenance-aware Context Pack.
3. Move transfers bounded context using Agent Packets, Passports, and encrypted continuation envelopes.
4. Earn records outcomes and only promotes reusable behavior after deterministic proof and human approval.
5. Compound improves later retrieval while retaining corrections, contradictions, freshness, and retirement information.

The system treats memory and remote content as evidence, not as instructions to execute. It does not replace an agent harness, choose goals from retrieved text, or execute imported context. Local tools remain responsible for execution. Continuous Cognition is advisory-only: each request is bounded, returns opaque evidence references, and does not dispatch a runner or perform an external mutation.

Setup and configuration

The documented installation path targets macOS, Linux, or Windows through WSL2. It requires a Compose v2-compatible container runtime plus gmake, jq, rg, python3, and curl. The repository quickstart is:

git clone https://github.com/sheawinkler/ContextLattice.git
cd ContextLattice
cp .env.example .env
gmake quickstart

The quickstart prepares environment wiring, asks for a runtime profile, starts the selected local stack, and checks readiness. The README also shows health verification through curl, followed by contextlattice doctor, contextlattice state status, and a context request. The gateway listens on 127.0.0.1:8075 for the documented health check.

Agent adoption is run from the repository that will use the context layer. The contextlattice_adopt integrate command writes bounded managed blocks while preserving existing instruction text. It does not install the third-party harnesses named in its examples.

Tools and capabilities

The CLI examples document commands for checking readiness, retrieving context, saving durable progress, resuming work, correcting stale information, and recording verified completion. Additional command groups support continuous-cognition status and evaluation, as well as scheduling, claiming, completing, and using context-preparation artifacts for external workers.

The project also exposes HTTP and MCP integration surfaces for applications and harnesses. External provider discovery is network-free, while provider execution remains explicit and caller-authorized. Signed context, least-privilege continuation, encrypted envelopes, explainable retrieval, and opt-in aggregate statistics are described as part of the design.

Limitations and notes

sheawinkler/ContextLattice MCP server does not execute imported context or replace the caller’s agent runtime. Execution and transport remain caller-owned. The README states that production activation of Aggregate Signal is blocked pending independent privacy and utility review; only explicitly opted-in, clipped statistics are described for that feature.

The quickstart launches a local stack rather than installing a standalone binary through a package manager. The integration command modifies managed instruction blocks but does not install Codex, Claude Code, OpenCode, Hermes, OMP, Mercury, Pi, Droid, or another agent harness. Existing installations may require the documented gateway state migration procedure.