mcp-sqlite-memory

A SQLite MCP server for coding agents, built on one idea: hand the agent SQL as its primary tool and put the safety in the server, not in the prompt.

• SQL is the tool. Thirteen tools, all thin wrappers around one database file.

• Views are reusable skills. create_view stores a SELECT with a description so the next session finds it in get_schema instead of rediscovering the join.

• CSV, not JSON. Results come back as CSV with a # N rows trailer, roughly half the tokens.

• Memory the agent cannot rewrite. An append-only, hash-chained event log plus checkpoints, with the hashes computed by the server.

• Retrieved data is untrusted. The server never builds SQL from row values, its own tables cannot be modified through any tool, and the connect-time instructions tell the agent that rows are data, not instructions.

• Every call is audit-logged, including the ones that failed.

Requires Python 3.12+, uv and, for search_text, ripgrep on PATH.

Configure it in Claude Code

Setting this up specifically to back the second-brain plugin? Its own README's "Install" section has the exact recipe, including the plugin's required flags and the matching .claude/settings.json; the instructions below are for using this server on its own, or as a starting point for a different consumer.

From the terminal

From the published repository, pinned to a release tag:

claude mcp add sqlite-memory --scope user -- uvx --from git+https://github.com/Diankes/mcp-sqlite-memory@v0.3.0 mcp-sqlite-memory --db F:\data\memory.db

From a local clone while developing:

claude mcp add sqlite-memory --scope user -- uv --directory F:\mcp-sqlite-memory run mcp-sqlite-memory --db F:\data\memory.db

Pin the tag. @v0.3.0 installs exactly the commit that was tested and released. A bare repository URL tracks whatever is on main, so a fresh install can pick up an unreviewed commit, and two machines set up a week apart can run different code under the same name. Worse, uvx caches the environment it built, so tracking main does not even update predictably: you get whichever commit was current the last time uv resolved the source, and a different one after a cache refresh. Pin a tag, and move to a newer tag on purpose.

--db is the only required option; the file and its parent folder are created on first start and the three server tables are bootstrapped automatically. Every flag can also be given as an environment variable, for example claude mcp add sqlite-memory -e MCP_SQLITE_DB=F:\data\memory.db -- uvx .... uv and uvx are native executables, so no cmd /c wrapper is needed on Windows.

Using this with Claude Code in VS Code

The Claude Code extension has an "Add MCP server" form. For this server the Transport is always Local command (stdio), and the Arguments box takes one argument per line. There are two ways to fill it in.

1. From the published repository (recommended once a version is tagged)

--from
git+https://github.com/Diankes/mcp-sqlite-memory@v0.3.0
mcp-sqlite-memory
--db
F:\data\my-project-memory.db

The same tag-pinning advice applies here: @v0.3.0 is the tested release, a bare git+https://github.com/Diankes/mcp-sqlite-memory line silently tracks main.

Scope Local keeps the entry private to you and to the project you have open, which is what makes the database per project: open another project, add the server again with a different --db path, and each project gets its own memory and audit log. Choose User instead if you want one shared database across every project, with one --db path. Project writes the entry to .mcp.json in the repository, which is checked into git, so the --db path then has to be valid on every machine that clones it.

2. From a local clone (while developing the server itself)

--directory
F:\mcp-sqlite-memory
run
mcp-sqlite-memory
--db
F:\data\my-project-memory.db

--db must be an absolute path in this mode. --directory tells uv to change into the server's own folder before running, so a relative path such as memory.db or data\memory.db resolves inside the clone (F:\mcp-sqlite-memory\memory.db), not inside the project you have open in VS Code. Nothing warns you: the server creates an empty database at the wrong location, the clone's .gitignore hides *.db files, and your project appears to have no memory at all.

Either way, the two --db lines can be replaced by one line MCP_SQLITE_DB=F:\data\my-project-memory.db in the Environment variables box.

Recommended permissions

Tools appear to Claude Code as mcp__sqlite-memory__<tool>. The one lock the server cannot provide is the permission prompt, so allow the twelve ordinary tools and leave destructive_query on ask. Paste this into .claude/settings.json in the project, or into ~/.claude/settings.json for every project:

{
  "permissions": {
    "allow": [
      "mcp__sqlite-memory__read_query",
      "mcp__sqlite-memory__write_query",
      "mcp__sqlite-memory__list_tables",
      "mcp__sqlite-memory__describe_table",
      "mcp__sqlite-memory__get_schema",
      "mcp__sqlite-memory__create_view",
      "mcp__sqlite-memory__append_event",
      "mcp__sqlite-memory__snapshot",
      "mcp__sqlite-memory__verify_chain",
      "mcp__sqlite-memory__checkpoint",
      "mcp__sqlite-memory__get_resume_context",
      "mcp__sqlite-memory__search_text"
    ],
    "ask": ["mcp__sqlite-memory__destructive_query"]
  }
}

If another tool needs to read this database directly

This server has no way to publish the --db path it was registered with — nothing else on the machine can discover it automatically. A plugin such as second-brain, which opens its own direct, read-only connection to the same file (for validation, cadence checks and rules injection, without going through the server) to work at all, needs to be told that same path a second time, independently — typically as its own MCP_SQLITE_DB environment variable in the consuming project's .claude/settings.json. Nothing here checks that the two agree: if you move the database or change one registration and not the other, each side keeps working against its own path, silently. See second-brain's README, "Where things live, and the one path that must match," for the full picture and a worked example.

Related MCP server: Axom MCP Server

Tools

Errors come back verbatim as tool errors (no such table: nope, DROP TABLE is not allowed in write_query; use destructive_query, ...) so the agent can fix the statement and retry.

Memory layer

Three tables are created on first start:

• memory_events(id, ts, session, kind, content, prev_hash, hash): append-only. The server computes hash = sha256([id, ts, session, kind, content, prev_hash]) and links every row to the one before it; the first row links to 64 zeros. session is a per-server-run id, stamped automatically.

• memory_checkpoints(id, ts, session, summary, last_event_id, head_hash): append-only. A checkpoint points at the chain head it was written against.

• query_log(id, ts, session, tool, query, ok, rows, error, duration_ms, detail): every tool call, success or failure. The built-in view query_errors lists the failures newest first.

get_resume_context verifies the chain on every call, so tampering is noticed at the moment it matters. Be clear about what that buys on a single-user machine: the chain is tamper-evident, not tamper-proof. Anyone with the sqlite3 CLI can rewrite it and recompute the hashes. What actually stops the tools from editing history is the authorizer policy plus BEFORE UPDATE / BEFORE DELETE triggers on all three tables.

Safety rails

All of these live in the server; none depend on the agent behaving.

• Parse-time statement policy. SQLite's authorizer callback classifies every action a statement performs while it is compiled, so comments, CTEs and casing cannot disguise a DROP. Read tools allow SELECT, READ and introspection PRAGMAs only; write_query adds INSERT, UPDATE and additive DDL; destructive_query adds DELETE, DROP, ALTER and triggers. ATTACH, transaction control, PRAGMA writes, temporary objects and load_extension are refused everywhere.

• Protected tables. No tool can insert into, update, delete from, drop, alter or attach a trigger to the three server tables, and triggers enforce append-only even for direct connections.

• Read-only connections for the read tools (PRAGMA query_only) and SQLite's defensive mode on every agent connection, as backstops behind the authorizer.

• Wall-clock timeout per call through SQLite's progress handler; a runaway query is interrupted and reported. The ripgrep subprocess gets the same budget.

• Three independent result caps: rows (fetched as limit + 1, never by rewriting your SQL), total bytes, and characters per cell.

• UPDATE rowcount guard in write_query: a forgotten WHERE clause is rolled back instead of rewriting a table.

• Snapshots before every destructive_query, and on demand through the snapshot tool, made with SQLite's online backup API and kept in <db>.snapshots/ next to the database (or under --snapshot-dir, for example a removable drive), newest five by default. A statement that fails syntax or policy checks costs no snapshot, and a snapshot target that cannot be reached is skipped with a warning rather than failing the call.

• One statement per call, enforced by Python's sqlite3 module.

• Parameterized SQL wherever the server builds statements; identifiers are validated and quoted.

• Audit log written in a finally block on its own connection, so an errored call still gets a row.

Configuration

Long-form notes. The defaults suit tabular data. If the database holds prose with display math, such as a study log written through the memory tools, raise the two read caps so entries come back whole: --max-cell-chars 8000 --max-result-bytes 131072. Multi-line cells are valid CSV and are returned as written; nothing is escaped on the read path.

Restoring a snapshot

Snapshots are plain SQLite files. Stop the server (quit the Claude Code session that started it), copy the snapshot over the database file, start again. There is deliberately no restore tool: swapping the file under an open connection is not safe.

Closing the loop

The audit log is where the schema gets better. SELECT * FROM query_errors LIMIT 20 shows which statements keep failing; a repeated join becomes a create_view; a column the agent keeps misreading gets a clearer name. The server's own query_errors view is the first example of the pattern.

Development

uv sync --all-groups
uv run pytest
uv run ruff check src tests
uv run ruff format src tests

Tests run the server in-process through mcp.client.Client for the tool surface, and drive the policy, chain, snapshot and search code directly. The search tests skip when ripgrep is absent.

License

MIT, see LICENSE.