Passing executions
Local run on November 10, 2026, commit aec177d0. Zero failures, 26 ignored. Includes the CLI suite repeated for both executables; not a count of unique tests.
Lodeward / governance system
Memory for AI.
Human control.
Connect context across sessions. Check commands before execution. Review what your agent learns before a lesson becomes a rule.
Explore the system ↗RUST / MCP / HOOKS
LOCAL BY ARCHITECTURE
LOCAL BY ARCHITECTURE
Brain / 3D synapses
Conceptual representation
Conceptual representation
01 / INTENT · RULE · DECISION · MEMORYN1 · under validation
Private repository
Private repository
01 / The purpose
Autonomy needs
visible boundaries.
Lodeward connects persistent memory and execution rules to the AI client. The MCP server handles queries; hooks act at workflow checkpoints. A learned lesson becomes guidance only after your approval.
02 / Workflow
Every stage has
an owner.
Select a stage.
Every decision is readable, even without 3D.
A gate before the command.
A hook checks command patterns before execution. Policy can allow, request confirmation or block. The detector is lexical: it cannot recognize every attack and does not replace a sandbox.
01 command received
02 policy checked
03 allow / confirm / block
Learning is not permission.
A recorded lesson enters quarantine. It should not guide the agent while awaiting review. Graph observations and approved lessons follow different paths.
01 lesson recorded
02 status: pending
03 use as a rule: not authorized
The rule is still yours.
You review the lesson and choose to approve or reject it. Change proposals have a separate cycle: approval, application and rollback. This page does not simulate human authorization.
01 owner review
02 approve or reject
03 decision recorded
Context that crosses sessions.
The MCP client queries the graph and approved lessons. Hooks deliver selected context; the compaction archive preserves material for retrieval. Better retrieval and lower usage remain hypotheses to measure.
01 MCP client query
02 memory selected
03 context with provenance
Explanatory demonstration. It does not execute commands, access your memory or approve lessons. Real integration depends on hooks installed in the client.
Interface / actual capture
The workflow,
made visible.
Input. Execution. Review. Memory.
See relationships and decision points in the integrated harness Agents page.
03 / Evidence
Proof before
the promise.
Historical record
November 9–10, 2026
Scope stated for every result
Compatible catalogs
tools/list matched the captured 70e415e8 baseline in eight scenarios. Schemas, descriptions and array order were preserved. The baseline was not rerun in that round.
Lesson workflow cases
Quarantine, approval, rejection, restart and project filters in an isolated demonstration with synthetic data. This does not prove universal isolation or quality gains.
Prototype, by the rubric
Assessment of November 9, 2026. Reproducible installation passed; the full V1–V8 rubric still has gaps. Rename validation does not promote the product to N3.
These figures belong to the stated commit and protocol. They do not validate the latest version. Read sources and hashes ↗
04 / Project status
The limits are
part of the work.
The engine runs and the lesson cycle has been demonstrated. General retrieval superiority, long-session savings and public readiness remain unproven.
- 01Live security checks and the complete interface test suite.
- 02Compaction A/B comparison and long-session evaluation.
- 03Adversarial review and an audit of every claim.
- 04Fourteen days of real usage before claiming daily readiness.
What earlier measurements did not demonstrate
Retrieval: 457 versus 456 in 470 queries, interpreted as a tie with BM25. Long sessions: ratio 1.2133 (95% CI 1.1039–1.3397), about 21% more usage in 45 pairs. Reduced context: 30 of 40 answers versus 40 of 40 with full context. Model tests used partial isolation. These are historical results, not rerun by this page.
Lodeward and the previous name
Lodeward continues the SynapseEngine project. Compatibility preserves the synapse alias and existing MCP contracts. A new identity does not mean every technical gap has been resolved.
Visual direction and references
Organic form and brain data visualization: Brain.space / TOMMY&YOU — Awwwards, including cursor interaction and the technology page composition. The studio case study connects human form with precise data. Rendered in Three.js with original illustrative connections. Brain model: BodyParts3D, © The Database Center for Life Science, under CC BY 4.0, converted by OrganUI and adapted for this page. Credits and modifications. Synapses represent software stages, not live data.