@nanpaidashi/dsh-honcho-sync

Honcho Memory Plugin for DeepSeek Harness — give your AI persistent, cross-session memory backed by a self-hosted Honcho v3 service.

中文 | English

---

What it does

• Auto-sync: Every conversation turn is automatically synced to Honcho (with debounce). Sync cursors persist across restarts.
• Hybrid recall: Semantic search across both raw messages AND reasoned conclusions (explicit/deductive/inductive) in parallel.
• Memory injection: On session start, relevant peer cards, representations, and semantic recall results are injected into the system prompt.
• 25 tools: Full Honcho v3 API surface exposed as DSH tools — search, profile, session management, peer management, conclusions, dreaming, queue monitoring.

Requirements

• DeepSeek Harness (DSH) with web profile
• A running Honcho v3 instance (self-hosted)
• Node.js >= 18

Installation

# From npm
dsh plugin --profile web add npm:@nanpaidashi/dsh-honcho-sync dsh web

# Or from GitHub
dsh plugin --profile web add link:https://github.com/nanpaidashi/dsh-honcho-sync dsh web

Then restart DSH:

systemctl --user restart dsh-web
# or: dsh web --restart

Upgrading

If you installed an earlier version (v0.6.x or v0.7.0):

# Re-run the same install command — it pulls the latest version and overwrites
dsh plugin --profile web add npm:@nanpaidashi/dsh-honcho-sync dsh web
# or
dsh plugin --profile web add link:https://github.com/nanpaidashi/dsh-honcho-sync dsh web

systemctl --user restart dsh-web

The install is idempotent: it replaces the plugin source file in place. Your configuration (settings panel) and sync state (~/.dsh/honcho-sync-state.json) are preserved — no data loss, no re-configuration needed.

v0.8.0 → v0.8.1 changelog

• Context efficiency: Reduced registered tools from 25 to 4 (honcho_recall, honcho_ask, honcho_remember, honcho_context) — saves ~7K tokens of tool schema injection per session
• Injection quality: Added _filterRepresentation() — removes Pattern [medium]/[low] blocks, Premises/Sources provenance chains, and orphan Type/Sources lines from representation injection
• Injection quality: Added _assembleByPriority() — assembles memory-context parts by priority (peer-card > session-summary > semantic-recall > representation) instead of fixed order
• Injection budget: injectionMaxChars default 8000 → 4000, reprMaxObs default 8 → 4
• Tool descriptions: Shortened all tool descriptions to ~100 chars

v0.7.x → v0.8.0 changelog

• DSH 0.1.0-rc.8 compatibility: Added normalizeToolParameters() to convert legacy parameter shorthand to standard JSON Schema
• DSH 0.1.0-rc.8 compatibility: Added output.render returning content blocks array (required by DSH >= 2026)
• README: Removed settings panel references — configuration is done via environment variables in ~/.config/systemd/user/dsh-web.service
• Fixed conclusions/query requiring filters parameter
• Fixed response parsing for /search, /peers/list, /summaries
• Fixed POST /conclusions body format
• Fixed schedule_dream parameter name (dream_type)
• Added 4 tools: honcho_conclude, honcho_conclude_list, honcho_conclude_query, honcho_message_get
• Hybrid recall now searches messages + conclusions in parallel

Uninstalling

…