Second Brain is a privately developed AI knowledge management system — designed to make 16+ years of professional knowledge searchable, reusable, and callable from any AI interface.
The system pairs a vector-embedded PostgreSQL database with an MCP server, making the entire knowledge base callable as a live tool across Claude Desktop (via mcp-remote bridge), OpenAI Codex CLI, and Microsoft Copilot Studio.
Retrieval is semantic: queries are embedded at call-time and matched against stored vectors via cosine similarity — surfacing related knowledge regardless of exact phrasing.
Knowledge is organized by source layer and tagged at ingestion. Each layer has distinct retrieval characteristics — dense technical standards behave differently than conversational memory or process documentation.
The knowledge layer is decoupled from any single AI vendor. Language models are interchangeable consumers; the data, access rules, and classification live in one place and travel with the user, not the tool.
Knowledge is exposed through the Model Context Protocol rather than a vendor plugin format. Embedding models are routed separately, so neither the client nor the embedder is hard-wired.
A single server function serves every client, so retrieval behaviour and classification rules are identical regardless of which model asks.
Each device authenticates with its own key and is bound to an access profile that limits which tools it can call and which knowledge bases it can read. Newly created knowledge is private to the owner until explicitly released.
Devices and knowledge bases are registered, so either can be deactivated or archived independently. Access ends immediately and shared secrets never need rotating.
Several deliberate constraints shaped this system from the start — they're not features, they're load-bearing decisions.
Client data is never ingested. Source tagging is enforced at ingestion time, not as a post-hoc audit. Supabase is pinned to Canada East for data sovereignty. Row Level Security is active.
Knowledge is structured as "Playbooks" — executable knowledge units that define context, inputs, procedure, and output independently of any specific AI interface or client.
Third-party content (e.g. external documents) is ingested only on manual trigger, with explicit metadata tagging. No passive crawling; no ambiguous provenance.
A unified source taxonomy was applied across all ~19,000 rows via SQL migration. Retrieval can be filtered and audited by source at any granularity.
The system is operational and actively used as a live tool across Claude Desktop, OpenAI Codex CLI, and Microsoft Copilot Studio. Core ingestion pipelines are complete. Per-device access control is deployed, and portability is being extended across additional LLM platforms. Moving to deeper contextual classification, to capture and describe thoughts by the context it was learned in, who it concerns, where it belongs, and how sensitive it is, so ‘who may see this, and why?’ has a precise answer. This enables speed and depth of semantic retrieval, while securing decisions around access based on deeper meaning and provenance.
The architecture has undergone a formal security review covering credential management, secret rotation, data sovereignty, audit logging, and client data exclusion — with remediations documented and applied.