Core concepts
Oka is built around a small set of ideas. You can use the product without reading this page, but a few minutes here will save you hours later.
1. Friction events
Traditional chat assistants forget everything between sessions. Oka’s Reason product captures structured friction events — decisions, deviations, observations, completions, learnings — emitted by agents and humans as they work.
These events are the raw signal. Every block, workaround, retry and escalation is a data point. Individually they’re noise; in aggregate they reveal where your organization is systematically stuck.
2. Consolidation
A background process periodically reviews raw events and distils them into learnings — structured insights with confidence scores, tags and source attribution. This isn’t summarization; it’s intelligence synthesis:
- Contradictions are resolved (“Team initially chose X, migrated to Y after discovering Z — Y is now canonical, confidence 0.85”)
- Patterns that span sessions and teams are surfaced
- Old learnings lose confidence if new evidence contradicts them
The consolidation engine is the core technical differentiator. It determines
whether context("auth") returns noise or genuinely useful institutional
knowledge.
3. Organizational intelligence
When an agent calls context("auth"), it gets compound understanding from
hundreds of prior sessions — not a raw dump, but refined learnings ranked by
confidence. The unit of analysis is the organization, not the individual
agent.
This is what distinguishes Oka from per-agent memory or RAG-over-codebase tools. Individual agent sessions are ephemeral. Organizational intelligence compounds.
4. MCP (Model Context Protocol)
Reason speaks MCP, an open protocol for exposing tools and context to LLM clients. Any MCP-compatible client (Claude Desktop, Claude Code, Cursor, Codex, Aider, etc.) can connect to Reason and call its tools directly.
The MCP server is distributed as @oka-core/reason
on npm — a single npx command to connect.