Whitepaper
Why knowledge graphs compound where RAG plateaus.
Every time you ask an AI a question, it starts from scratch. Chat history evaporates. Meeting notes pile up unread. Strategic context built over months disappears when a tab closes. The AI is powerful at the moment of query - but it learns nothing, retains nothing, and compounds nothing. Your organisation's intelligence keeps leaking.
Retrieval-Augmented Generation is good at finding relevant text. It is not good at maintaining understanding. The same insight gets re-derived on every query. Contradictions accumulate silently. The corpus grows; the comprehension does not. Kernal takes the opposite bet: synthesise at write time, not at query time. Every source that enters the system updates a maintained knowledge base. The wiki gets better with every source added - not just longer.
Most AI products sit at the agent layer - they do things. Kernal sits underneath, as substrate - it is what the agent knows. The lineage is Palantir Foundry: master-data at the bottom, ontology in the middle, AI execution on top. We are not Palantir; we are deliberately the simpler, more portable bet, scoped to professional services rather than government and finance. The agents you buy or build sit on top of Kernal. The graph below them is what makes them useful, learnable, and yours.
Kernal structures your team's knowledge across five layers: Intent (goals, strategy), Commercial (deals, relationships), Intelligence (patterns, insights), Execution (actions, decisions), and Conversation (meetings, transcripts). Each layer has its own schema and relationships. Together they form a graph - one that can be traversed, queried semantically, and exported as a portable SQLite file. Served over MCP, it connects to Claude, ChatGPT, and any MCP-compatible AI tool.
Most "AI memory" products demand a stack - Postgres, Pinecone, Redis, an auth service, a queue, monitoring, a vendor commitment. Kernal is one SQLite file. Vector search runs inside it via the sqlite-vec extension. Embeddings come from a local model - no API call out, no data leaves your perimeter. Backup is copying a file. Migration is sqlite3 dump. The simplicity is not aesthetic; it is the only way a 20-person firm adopts this without it becoming a second IT department.
Most "AI for X" tools fail because they require someone to remember to feed them. Kernal installs lifecycle hooks into the agent host - Claude Code, Cursor, Codex, Copilot, any MCP-compatible runtime - that fire at four key moments. Session start: pre-load the relevant slice of the graph. Every prompt: route context as you type. Before context-window compaction: compress long-term memory before short-term gets evicted. Session end: capture observations into a review queue. The hooks are why an agent goes from "smart chatbot" to "actual second brain." Memory happens whether anyone is paying attention or not.
Every enterprise has shadow AI. People bring their own ChatGPT, their own Claude, their own Copilot to work because the corporate tool requires too much manual data entry, has wrong defaults, and changes faster than priorities. Kernal does not compete with that. Kernal is the substrate underneath whichever AI host the user already trusts - same principle as Outlook or shared drive: a company asset issued on day one, surfaced through tools the user already uses, captured automatically as a byproduct of work that would happen anyway. The individual incentive is less admin: less manual CRM updating, less context-rebuilding, less coordination overhead. Knowledge accumulation is the side effect. That is why bottom-up adoption survives where top-down rollouts die.
In production use: 8 clusters analysed in a single batch, 8 cluster meta-pages written automatically, 3 cross-cluster contradictions surfaced - including a critical tension between a $170B market bet and the same company's missed revenue targets. No human read across all 50 source pages. The system did. In professional-services pilots, the cost-of-coordination math typically pays back within the first delivered engagement. The difference is architectural. Capability is a ceiling. Context is a compounding asset.
The same write-time synthesis that makes Kernal useful also makes it accountable. Because every source, decision, and synthesis step is recorded as it happens rather than reconstructed after the fact, the system accumulates attributable decision records and audit trails as an ordinary byproduct of doing the work. What was the input context, what did the model produce, who reviewed it, and when: those questions already have structured answers, because answering them is how the substrate maintains itself. This is an architectural insight, not a feature you switch on. A system that synthesises at write time cannot help but leave a legible trail behind it. That trail does not meet an organisation's obligations on its own, and it makes no such guarantee. It produces the attributable records and audit trails that support your compliance programme, so that when a regulator asks what your AI saw and what a human decided, the answer is already structured rather than reassembled under pressure.
Capability is a ceiling.
Context is a compounding asset.