How It Works

Cogni-Code runs a continuous loop with four phases: capture, process, inject, evolve.

  your conversation
        │
        ▼
  ┌──────────────┐     ┌───────────────┐
  │ session hooks │────▶│ graph_memory  │   ← MCP tool you call directly
  │ capture state │     │ tool surface  │     (remember, recall, search...)
  └──────────────┘     └──────┬────────┘
                              │
                 ┌────────────▼────────────┐
                 │     background pipeline  │
                 │                          │
                 │  scribe → auditor →      │   ← active pipeline (always runs)
                 │  librarian → dreamer     │
                 │                          │
                 │  observer (writes to     │   ← always active, single node store
                 │  nodes/ alongside main)  │
                 └────────────┬────────────┘
                              │
                 ┌────────────▼────────────┐
                 │     ~/.graph-memory/     │
                 │                          │
                 │  mind/model.json         │   ← cognitive profile (always active)
                 │  lenses/{project}/       │   ← project models (always active)
                 │  sessions/{project}.jsonl│   ← session logs (always active)
                 │  nodes/                  │   ← knowledge graph (canonical store)
                 │  dreams/                 │   ← creative associations
                 └─────────────────────────┘

Capture

Session hooks watch your conversations and tool traces. Nothing is sent anywhere. The hooks write to a local buffer.

Process

The scribe extracts structured deltas from the buffer. The auditor detects stale and contradictory nodes. The librarian applies judgment-heavy updates with a prune-over-preserve philosophy. The dreamer creates speculative cross-node associations. None of this is your agent trying to manage its own memory — it is a separate pipeline that runs after the fact.

Inject

The next session starts with layered context: who you are (mental model), how this project works (project lens), what happened recently (session log), and what is most relevant (per-project knowledge index). The total cost is a few thousand tokens. Your agent resumes where the last one left off instead of reconstructing the world from scratch.

Evolve

Memory decays when unused. Nodes archive gracefully. Dreams surface unexpected connections. Skillforge promotes frequently-recalled memories into slash commands. Everything is git-backed and reversible.