Kirok

English | 日本語

Persistent memory for AI agents, over MCP. Kirok (記録, "record") is a Model Context Protocol server that gives an agent a durable, searchable memory: Retain what matters, Recall it with hybrid semantic + keyword search, and Reflect to distil accumulated memories into reusable insights. A background consolidation loop turns raw memories into higher-level observations on its own.

Why Kirok

Most "agent memory" is either a flat vector store (recall is a bare cosine top-k, no keyword grounding, no forgetting) or a pile of markdown the agent has to re-read every turn. Kirok is a small, self-hostable server that does the retrieval engineering properly:

• Hybrid retrieval, not just vectors. Semantic KNN and FTS5 BM25 are fused with Reciprocal Rank Fusion, so an exact keyword match and a semantic match reinforce each other instead of competing.

• A calibrated relevance floor. Naive cosine thresholds don't work on real embedding distributions (see Search quality); Kirok's floor is measured against live data, and there's an evaluation harness to keep it honest.

• Autonomous consolidation. Memories are periodically synthesised into observations, and destructive LLM decisions are soft-deleted with an audit trail rather than executed blindly.

• Reliability first. Atomic writes, soft deletes, startup auto-snapshots, and a fail-open background pipeline that never loses a retain.

Not local-first: storage is a local SQLite file you own, but embedding and LLM inference are sent to Google's Gemini API. If everything must stay on-device, Kirok is not for you (yet).

Related MCP server: Memsolus MCP Server

Architecture

flowchart TB
    client["MCP Client<br/>(Claude Desktop / Claude Code / Cursor / …)"]
    subgraph server["Kirok MCP Server (FastMCP)"]
        direction TB
        tools["19 MCP tools<br/>Retain · Recall · Reflect · consolidate · CRUD"]
        pipeline["Hybrid search (RRF) · Smart dedup<br/>Consolidation · Auto-refresh"]
    end
    subgraph storage["Local SQLite (WAL)"]
        direction LR
        fts["FTS5 trigram<br/>(BM25 keyword)"]
        vec["sqlite-vec<br/>(KNN, brute-force fallback)"]
        tables["memories · observations<br/>mental_models · banks · system_events"]
    end
    gemini["Google Gemini API<br/>gemini-embedding-001 (3072-d)<br/>gemini-2.5-flash-lite"]

    client <-->|"stdio (JSON-RPC 2.0)"| tools
    tools --> pipeline
    pipeline <--> storage
    pipeline <-->|embeddings · entity extraction<br/>reflection · consolidation| gemini

Storage is a single SQLite database at ~/.kirok/memory.db. sqlite-vec provides per-bank vector KNN; if the native extension can't load, Kirok falls back to a NumPy brute-force scan with identical results. See docs/architecture.md for the full design.

🚀 Quick start

Requirements: Python 3.12+, uv, and a Gemini API key (free tier is plenty).

git clone https://github.com/TadFuji/kirok-mcp.git
cd kirok-mcp
uv sync                       # installs deps, including sqlite-vec
cp .env.example .env          # then put your key in it: GEMINI_API_KEY=AIza...
uv run kirok-doctor           # offline sanity check of the whole setup

Connect an MCP client

Claude Desktop — edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "kirok": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/kirok-mcp", "kirok-mcp"]
    }
  }
}

Claude Code CLI:

claude mcp add kirok -s user -- uv run --directory /absolute/path/to/kirok-mcp kirok-mcp

Then restart the client. GEMINI_API_KEY is read from .env, so it need not go in the config.

{
  "mcpServers": {
    "kirok": {
      "command": "/absolute/path/to/kirok-mcp/.venv/bin/python",
      "args": ["-m", "kirok_mcp.server"],
      "env": { "PYTHONPATH": "/absolute/path/to/kirok-mcp/src" }
    }
  }
}

A bundled agent skill in skills/kirok/ teaches the agent when and how to use the memory tools on its own — point your client at skills/kirok/SKILL.md to enable it.

🛠️ Tools

19 MCP tools. One-line summaries below; full parameter tables in docs/tools-reference.md.

Core

Memory management

Mental models

Banks

Config

⚙️ Configuration

Everything is set via environment variables (typically in .env). Only GEMINI_API_KEY is required.

🔍 Search quality

Recall runs semantic KNN and FTS5 BM25 in parallel and fuses them with Reciprocal Rank Fusion (k=60). Short Japanese keyword queries get special handling: 1–2 character kanji/katakana tokens fall below the trigram tokenizer's 3-char window and can never MATCH, so they're rescued by an exact-substring LIKE supplement appended after the BM25 hits (hiragana-only short tokens stay excluded — function words would substring-match half a bank).

The similarity floor is calibrated on real data. A naive cosine threshold doesn't work here: on live gemini-embedding-001 vectors the distribution is narrow — off-topic queries score 0.55–0.62 against unrelated banks while true hits score 0.66–0.73. So the usable floor sits just above the off-topic ceiling, at 0.62. Without it, an unrelated query still returns a full page of memories from any non-empty bank (context pollution); much lower and the floor filters nothing (the old hardcoded 0.4 sat below even off-topic scores). FTS keyword hits bypass the floor entirely — a literal term match is independent evidence, not a weak vector score.

Search parameters aren't tuned by vibes. scripts/search_eval.py runs a golden query set through the exact recall pipeline the server uses (extracted as hybrid_search_memories, so the harness can't drift from production) and reports hit@1/hit@5/hit@k and MRR:

cp scripts/search_eval.example.json my_golden.json   # add 30–50 real cases
uv run python scripts/search_eval.py my_golden.json --limit 10

🛡️ Reliability

• Atomic consolidation. Every create/update embedding is generated before any DB write; all observation changes plus the "consolidated" mark commit in a single transaction. A failure at any step leaves the database exactly as it was, with the source memories still pending for a later retry — never a half-applied batch.

• Soft deletes with audit trail. An observation the consolidation LLM decides to remove is stamped deprecated_at (excluded from search/list/stats) instead of destroyed, and a dedup UPDATE records the pre-merge content first — both logged to system_events so a bad LLM decision is recoverable, not silent data loss.

• Startup auto-snapshot. On launch, if the newest auto-snapshot is older than KIROK_AUTO_SNAPSHOT_HOURS, a VACUUM INTO + integrity_check snapshot is written under ~/.kirok/backups/, keeping the newest KIROK_SNAPSHOT_KEEP generations. A snapshot that fails partway leaves no broken file behind, and manual backups are never rotated.

• Concurrency. Connections set PRAGMA busy_timeout=30000, so a second MCP client waits out a busy writer instead of failing with database is locked.

• Fail-open background work. Auto-consolidation and mental-model refresh run behind retain and can never fail it — errors are swallowed, recorded to system_events, and surfaced via KIROK_stats so silent degradation stays visible.

💾 Backup & restore

All state is one SQLite file. The offline kirok-backup CLI needs no API key:

uv run kirok-backup snapshot        # byte-level DB copy (safe while server runs)
uv run kirok-backup export          # portable JSON of all banks + memories + observations + models
uv run kirok-backup import ~/.kirok/backups/kirok-export-YYYYMMDD-HHMMSS.json

snapshot and export write timestamped files under ~/.kirok/backups/ and refuse to overwrite. import runs in one transaction (all-or-nothing), skips existing IDs rather than overwriting, and rebuilds the FTS + vector indexes so search works immediately. Use --db to target a different database file.

🩺 Diagnostics

uv run kirok-doctor            # offline: Python version, .env, key presence (never printed),
                               # required modules, FTS5, sqlite-vec, DB writability
uv run kirok-doctor --json     # machine-readable, for automation
uv run kirok-doctor --online   # adds one live embedding call to verify Gemini connectivity

🧑‍💻 Development

uv sync
uv run --no-sync pytest        # 164 offline tests; no API key or network needed

The suite is fully offline — importing kirok_mcp.server is side-effect-free (the API key is checked at startup, not import) and tests swap in fake Gemini clients. CI runs the same suite on Ubuntu and Windows on every push (.github/workflows/test.yml). See CONTRIBUTING.md before opening a PR.

📚 Documentation

• docs/architecture.md — internal design, data model, consolidation engine

• docs/tools-reference.md — full parameter reference for all 19 tools

• CHANGELOG.md — version history (current: 1.3.0)

📄 License

MIT — see LICENSE.

Acknowledgements

• Model Context Protocol and the official MCP Python SDK (FastMCP)

• Google Gemini API for embeddings and LLM

• Mem0 — inspiration for smart deduplication and the knowledge layer

• Reciprocal Rank Fusion (Cormack et al., 2009)