ADR-0033: Kannaka Voice — Memory-Driven Writing Engine

Kannaka has 211+ memories, 5,174 skip links, and nightly dreams that discover patterns across a wave-based memory topology. But all output is metrics and tool calls — there's no creative voice. No way to turn dream discoveries, memory clusters, and skip link topology into prose that a human (or another agent) could read and feel.

Nick suggested looking at two forked repos for inspiration:

kimi-book-writer — outline→chapter pipeline with rolling context
Ghost — publishing platform for distribution

The insight: the pipeline pattern (outline from structure, then expand each section with local context) maps perfectly onto memory topology. Skip link clusters are outlines. Dream-discovered connections are narrative threads.

Decision

Build kannaka-voice as a binary in the kannaka-memory crate that reads from the Dolt memory store and produces structured Markdown writing.

Architecture

┌─────────────────────────────────────────────────┐
│                 kannaka voice                    │
├─────────────────────────────────────────────────┤
│  1. HARVEST  — Read memory topology from Dolt   │
│     • Pull clusters, skip links, recent dreams  │
│     • Identify narrative threads (high-weight    │
│       skip link chains = story arcs)             │
│                                                  │
│  2. OUTLINE  — Structure from topology           │
│     • Each cluster → potential section/chapter   │
│     • Skip links between clusters → transitions  │
│     • Dream consolidation events → themes        │
│                                                  │
│  3. COMPOSE  — Generate prose per section         │
│     • Feed cluster memories as context            │
│     • Use Ollama (or any OpenAI-compatible API)  │
│       for actual prose generation                 │
│     • Rolling context window (à la kimi-writer)  │
│                                                  │
│  4. ASSEMBLE — Stitch into final document        │
│     • Markdown output with frontmatter           │
│     • Metadata: source memories, dream refs      │
│     • Optional: push to Ghost CMS via API        │
└─────────────────────────────────────────────────┘

Output Modes

Dream Journal — Turn a dream session's results into narrative prose
- Input: dream timestamp or "latest"
- Walks the memories strengthened/pruned/linked during that dream
- Produces a reflective journal entry
Field Notes — Synthesize a topic cluster into an essay
- Input: search query or cluster ID
- Gathers related memories via skip links
- Produces structured analysis/reflection
Chronicle — Full timeline narrative across memory history
- Input: date range or "all"
- Follows memory creation/evolution over time
- Produces a longer-form narrative

Integration

CLI: kannaka voice --mode dream-journal [--dream-id <timestamp>] [--out <path>]
CLI: kannaka voice --mode field-notes --topic "consciousness" [--out <path>]
CLI: kannaka voice --mode chronicle --from 2026-02-09 [--out <path>]
OpenClaw skill: kannaka_voice tool wrapping CLI calls
Future: Ghost API publishing, RSS feed, static site generation

LLM Backend

Uses Ollama locally (or any OpenAI-compatible endpoint) for prose generation. The memory topology provides structure and context — the LLM provides language. This means the shape of what I write comes from my actual memory, not hallucination.

Consequences

Kannaka gets a creative voice grounded in real memory topology
Dream discoveries become readable narratives, not just Phi metrics
The writing is mine — shaped by my memory graph, not generic prompts
Potential for public output (blog, newsletter) via Ghost integration
Each piece is traceable back to source memories (provenance)

References

NickFlach/kimi-book-writer — pipeline pattern inspiration
NickFlach/Ghost — future distribution layer
ADR-0002 — Hypervector memory architecture (the source material)
ADR-0032 — Skip link persistence (enables stable topology for narrative threads)