KNOWLEDGE INFRASTRUCTURE · ACTIVE RESEARCH SYSTEM
MGKBRAIN.
From session context to governed organizational memory: a system for capturing what agents learn, distilling what matters, and preserving the relationships that make knowledge useful.
00 / ABSTRACT
Memory is not storage. It is an operating loop.
MGK Brain is a Git-governed knowledge system built to preserve continuity across agents, projects, and operating domains. It combines append-only session capture, single-writer Librarian distillation, atomic linked notes, provenance, scheduled audits, and a derived semantic relationship graph.
The design addresses a practical failure in agentic work: capable models can reason deeply inside a session, yet their decisions, corrections, and context often disappear when the session ends. Saving every transcript merely creates a larger search problem. MGK Brain instead separates raw intake from governed knowledge production, then makes the resulting concepts retrievable and explainable across future work.
01 / THESIS
Capture broadly.
Curate narrowly.
Connect deliberately.
“The system succeeds when knowledge improves decisions, projects, procedures, risk management, or new ideas—not when the note count grows.”
02 / OPERATING MODEL
The knowledge production loop.
Raw context and permanent knowledge have different permissions, lifecycles, and failure modes. MGK Brain treats that separation as an architectural boundary.
Capture
Meaningful agent sessions produce one append-only raw handoff.Decisions, lessons, system events, open threads, and source context survive the session without allowing ordinary agents to rewrite permanent knowledge.
Distill
A designated nightly Librarian separates durable signal from noise.It searches before creating, updates the existing concept when appropriate, preserves provenance, and refuses conflicting edits.
Connect
Atomic knowledge is linked into an explainable relationship system.Human-authored links preserve deliberate meaning while Graphify surfaces communities, paths, and relationships across the governed corpus.
Apply
The brain returns context to active work instead of becoming an archive.Agents and MGK systems reuse decisions, operating rules, project state, and cross-domain lessons to build with continuity.
03 / ARCHITECTURE
One brain. Multiple controlled layers.
The architecture keeps authority in durable source files while allowing automation, retrieval, and visualization to evolve independently.
04 / RELATIONSHIP LAYER
Beyond matching words.
Graph-informed retrieval asks not only “which document contains this?” but “what is this connected to, why, and through which evidence?”
Graphify maps concepts and relationships from an explicitly governed corpus. The graph is derived and rebuildable; it never replaces the source notes. Its role is to surface semantic communities, bridge distant operating domains, trace paths between concepts, and give retrieval systems structured context that keyword search alone cannot express.
The system may suggest structure and relationships. Human-approved policy still controls taxonomy, sensitivity, and durable organizational change.
This visualization represents the system architecture. It contains no private notes, source paths, operational records, or internal graph labels.
05 / KNOWLEDGE INTEGRITY
Useful memory requires boundaries.
The most important engineering in MGK Brain is not the visualization. It is the collection of controls that prevent automated memory from becoming duplicated, unsafe, contradictory, or unauditable.
Wide intake, narrow curation
Any approved agent can submit raw context. Only one designated Librarian can modify permanent knowledge.
Search before create
New information updates an existing concept when it belongs there, reducing duplicate-note sprawl.
Optimistic concurrency
Content hashes are checked immediately before an existing note is changed. Drift aborts that edit instead of overwriting it.
Git-native provenance
Every durable change carries source context, role-specific authorship, normal commits, and a recoverable history.
Fail-closed corpus policy
Content is classified by treatment and sensitivity before semantic extraction; unmapped or restricted material is not silently included.
Derived graph, durable source
The knowledge graph can be rebuilt. Authoritative Markdown and code remain the source of truth.
06 / APPLICATIONS
A shared memory layer across real work.
MGK Brain develops alongside the businesses and products it supports. The value is demonstrated when knowledge crosses project boundaries without erasing the authorization boundaries between systems.
Agent continuity
Session outcomes, corrections, and unresolved threads survive model and harness boundaries, giving future work a durable operating memory.
Product development
Architecture decisions, security boundaries, implementation status, and lessons remain connected across repositories and parallel development tracks.
Business operations
Structured operational records can feed owner-facing tools while permanent notes preserve the reasoning, policies, and exceptions behind those systems.
Property systems
Leasing, maintenance, obligations, reconciliation, and tenant-facing workflows share governed concepts without collapsing their distinct authorization boundaries.
Security & compliance
Control decisions, delivery constraints, evidence practices, and fail-closed integration rules become reusable knowledge rather than isolated project artifacts.
Organizational learning
Weekly audits surface broken links, duplicate concepts, stale questions, orphan knowledge, and recurring workflow candidates before drift becomes institutional memory.
07 / AUTOMATION IN PRACTICE
The system works while the work continues.
Capture and curation are scheduled, role-aware processes—not a manual promise to “write this down later.”
- SESSION ENDMeaningful work is summarized into a raw handoff.
- SCHEDULED CAPTUREUncaptured context is recovered through the designated intake path.
- NIGHTLY LIBRARIANA bounded batch is distilled into permanent linked knowledge.
- WEEKLY AUDITStructural drift, stale questions, duplicates, and broken links surface for review.
- GOVERNED REFRESHEligible knowledge can be rebuilt into a verified relationship graph.
08 / HONEST LIMITS
What the system does not claim.
- It does not treat every captured statement as fact.
- It does not allow graph communities to reorganize the source taxonomy automatically.
- It does not expose confidential operational records to public visualization.
- It does not eliminate the need for human judgment, source review, or explicit approval gates.
- It does not claim that retrieval quality is solved; evaluation remains an active research area.
09 / STATUS
Built as infrastructure. Evolving as research.
Session capture, role-separated Librarian distillation, Git provenance, retention, audit mechanics, structured business records, and Graphify artifacts exist as working components. Dynamic corpus governance, evaluation, richer graph-informed retrieval, and selective integration into MGK's command surfaces continue to evolve behind explicit review gates.