Skip to content

Overview

Cognibrain is a self-hosted engineering memory layer that gives coding agents durable recall across sessions. It stores corrections, repo rules, action outcomes, connector events, and patch evidence — then returns compact, relevant context before each agent action.

The Problem

Modern coding agents (Codex, Cursor, Copilot, custom LLM-based tools) are stateless by default. Every session starts fresh:

  • A reviewer corrects an agent's approach → the agent makes the same mistake next time
  • A build fails for a known reason → the agent re-discovers the fix from scratch
  • Team conventions exist in scattered docs → agents miss them consistently

How Cognibrain Solves It

graph LR
    A[Agent starts task] --> B[Request context from Cognibrain]
    B --> C[Receive relevant memories]
    C --> D[Guard risky actions]
    D --> E[Execute work]
    E --> F[Record outcomes & corrections]
    F --> G[Memory persists for next session]
    G --> B

Cognibrain acts as a feedback loop between agent sessions:

  1. Before acting — the agent requests a context pack (corrections, conventions, past outcomes relevant to the current task)
  2. Before risky operations — the action guard warns or blocks known-bad patterns
  3. After completing work — the agent records what happened (patch evidence, corrections, new facts)
  4. Between sessions — dream cycles consolidate, deduplicate, and surface stale memories for operator review

Integration Surfaces

Cognibrain exposes four surfaces, each optimized for a different consumer:

Surface Best for Protocol
CLI Operators, CI/CD, shell scripts Text + JSON stdout
Harness CLI Any shell-capable agent or git hook JSON stdin/stdout
MCP MCP-native agents (Codex, Cursor) Model Context Protocol
SDK/HTTP Product integrations, dashboards, custom runtimes REST API / TypeScript / Python

All surfaces share the same underlying memory engine and can run against the same local daemon.

Architecture at a Glance

graph TB
    subgraph Agents
        A1[Codex]
        A2[Cursor]
        A3[Custom Agent]
    end

    subgraph Cognibrain
        MCP[MCP Server]
        CLI[CLI / Harness]
        API[HTTP API]
        Core[Memory Core]
        Store[(Storage)]
        Conn[Connectors]
    end

    subgraph External
        GH[GitHub]
        Jira[Jira]
        Slack[Slack]
        PG[(Postgres)]
    end

    A1 --> MCP
    A2 --> MCP
    A3 --> CLI
    MCP --> Core
    CLI --> Core
    API --> Core
    Core --> Store
    Core --> Conn
    Conn --> GH
    Conn --> Jira
    Conn --> Slack
    Store --> PG

Key Concepts

Concept Meaning
Memory A durable, scoped fact that can be recalled before an agent acts
Operator The human responsible for inspecting and maintaining memory state
Connector An integration source/sink that ingests external events or writes back context
Harness An agent lifecycle integration that calls Cognibrain for context, guards, and outcomes
Dream Cycle A maintenance pass that detects stale memories and schedules operator review
Action Guard A pre-action check that warns or blocks known-bad operations
Patch Evidence A record of files changed, commands run, and memories used during a task

Open Source + Commercial

Cognibrain follows an open core model:

  • ✅ MIT (open source) — CLI, API, SDK, connectors, harness templates, MCP server, all documentation
  • 🔒 Commercial add-on — Operator UI (browser-based dashboard for visual memory management)

The commercial Operator UI is never required. Every feature is accessible through the CLI and API.