kiseki

Kiseki (軌跡) — Personal Memory for AI Assistants

Your AI forgets everything between sessions. Kiseki fixes that.

One Go binary. One SQLite file. 100% local. Zero cloud cost.

How It Works

                        ┌─────────────────┐
  Your notes, docs,     │                 │     Claude Code
  chat sessions    ───► │ Kiseki (SQLite) │ ◄── OpenCode
                        │                 │     Any MCP client
                        └────────┬────────┘
                                 │
                          Search Pipeline
                                 │
              ┌──────────────────┼──────────────────┐
              ▼                  ▼                   ▼
         FTS5 (exact)    Vec Chunks (semantic)  Vec Messages
              │                  │                   │
              └──────────────────┼───────────────────┘
                                 ▼
                     Merge + Tier Rank
               stones > solutions > key > normal

Queries go through entity expansion first — searching "Customer Guiding System" also finds "CGS" and related aliases — then hit three search layers in parallel. Results are merged, deduplicated, and ranked by importance tier.

Quick Start

# 1. Get Kiseki (fts5 tag required for full-text search)
go install -tags "fts5" github.com/Gsirawan/kiseki@latest
# Or download from Releases: https://github.com/Gsirawan/kiseki/releases

# 2. Get an embedding model (Ollama required)
ollama pull qwen3-embedding:0.6b

# 3. Use it
kiseki ingest --file notes.md
kiseki search "authentication setup"
kiseki status

Features

Commands

kiseki ingest       Ingest files (markdown, text, stdin)
kiseki search       Multi-layer search (entity + FTS5 + vectors)
kiseki search-msg   Search messages with context window
kiseki history      Chronological entity mentions
kiseki status       Health check + stats
kiseki serve        Start MCP server
kiseki watch-oc     Realtime messages ingesting - OpenCode session
kiseki watch-cc     Realtime messages ingesting - Claude Code session
kiseki batch-oc     Batch export OpenCode sessions - format (USER, time, date, message)
kiseki batch-cc     Batch export Claude Code sessions - format (USER, time, date, message)
kiseki stone        Manage solution/decision records
kiseki mark         Set chunk importance
kiseki review       Weekly review scoring
kiseki report       Generate weekly report
kiseki re-embed     Re-embed all vectors (model changes)
kiseki migrate      Import from sessions from DB
kiseki list         Show ingested files
kiseki forget       Remove chunks

MCP Integration

Add to .mcp.json (Claude Code) or opencode.json:

{
  "mcpServers": {
    "memory": {
      "command": "/path/to/kiseki",
      "args": ["serve"],
      "env": {
        "KISEKI_DB": "/path/to/memory.db",
        "KISEKI_PREFIX": "memory",
        "EMBED_MODEL": "qwen3-embedding:0.6b"
      }
    }
  }
}

KISEKI_PREFIX controls tool names — set it to nectar and tools become nectar_search, nectar_ingest, etc. Run multiple instances with different prefixes.

Entity Graph

Define entities in YAML, point to it with KISEKI_ENTITIES:

entities:
  - name: React
    aliases: [ReactJS, react.js]

  - name: Next.js
    aliases: [nextjs, next]

Searching "ReactJS" now expands to "React" across all search layers. Aliases are matched with word boundaries to prevent false positives.

Note Extraction

Auto-save subagent synthesis outputs as markdown files. The watcher detects completed subagent sessions, extracts the final response, and saves it as a structured note — zero manual intervention.

# Live: extract notes while watching a session
kiseki watch-oc --notes ~/notes --project-dir /path/to/project --notes-agent researcher

# Backfill: extract all existing sessions and exit
kiseki watch-oc --notes ~/notes --project-dir /path/to/project --notes-agent researcher --backfill

Notes are saved as {topic-slug}.md with metadata headers (Date, Agent, Session, Parent, Prompt). Duplicate sessions are tracked in .extracted and skipped on re-runs. Responses under 500 characters are filtered as non-substantial.

Configuration

Also supports: KISEKI_MAIL_IDENTITY, KISEKI_MAIL_PEERS, KISEKI_MAIL_DIR, KISEKI_TYPOS, KISEKI_BACKUP_DIR, KISEKI_NOTES_DIR. See .env.example.

Build

git clone https://github.com/Gsirawan/kiseki.git && cd kiseki
go build -tags "fts5" -o kiseki .

Stats

• ~15K lines of Go across 18 internal packages
• 6 direct dependencies (sqlite-vec, lipgloss, misspell, godotenv, go-sqlite3, go-sdk)
• Go 1.23+, Ollama required

License

MIT