mcp-memory-rs

mcp-memory-rs

Local-first MCP memory server in pure Rust. Persistent, versioned, queryable memory for AI agents — owned by the user, not by a platform.

State lives on your disk as plain JSON categories backed by SQLite (FTS5). Every write is versioned and backed up. A fleet of devices syncs through an optional HTTP endpoint with merkle manifests and optimistic concurrency. No cloud account, no embedding service, no network dependency in the default mode.

Why

Agent memory tied to a vendor dies with the vendor session. This server keeps the agent's long-lived state — identity, projects, infrastructure notes, per-device workflows — in files you can read, grep, diff, and back up yourself. Any MCP client (Claude Code, Codex, or anything speaking MCP stdio) gets the same memory; swapping the model does not lose the state.

Design choices that follow from that:

• Local-first: the stdio server works fully offline against local storage. Sync is a separate, explicit step — never a hidden dependency.
• Plain storage: one JSON file per category plus a SQLite index. The database can be rebuilt from the files; the files are the truth.
• Versioning by default: every write keeps history; deletes create a backup first. memory_history and memory_compact manage the tail.
• Concurrency-safe writes: expected_hash turns clobbering races between concurrent agents into explicit conflicts.
• Zero ML in the core: full-text search is SQLite FTS5 with BM25; memory_search_semantic adds an optional TF-IDF hybrid. No model downloads.

The pair: for *corpus* recall (chunked documents, BM25 retrieval,

original-text injection) see the companion server

mcp-vl-msa-rs. This server holds

the curated agent state; that one holds the queryable corpus.

flowchart LR

A["AI agent (any MCP client)"] A -->|"curated state read / write / sync"| M["mcp-memory-rs the notebook "] A -->|"corpus recall index / search / fetch"| V["mcp-vl-msa-rs the library "] M --- D1[("JSON categories SQLite FTS5")] V --- D2[("tantivy BM25 collections")]

Install

Prebuilt binary (recommended) — download the archive for your platform from the latest release, extract, and point your MCP client at the binary:

tar xzf mcp-memory-rs-x86_64-unknown-linux-gnu.tar.gz

install -m755 mcp-memory-rs-*/mcp-memory-rs ~/.local/bin/

Prebuilt targets (Linux + Android): x86_64-unknown-linux-gnu, x86_64-unknown-linux-musl, aarch64-unknown-linux-gnu, aarch64-unknown-linux-musl (edge / ARM / Termux), aarch64-linux-android.

macOS: no prebuilt binary is shipped (it would need Apple code-signing). Install from source instead — cargo install compiles it on your Mac in one command, no signing needed:

cargo install --git https://github.com/DioNanos/mcp-memory-rs --locked

--locked uses the committed Cargo.lock (reproducible build).

Quick start

cargo build --release

stdio MCP server, offline mode (default)

./target/release/mcp-memory-rs

Claude Code (~/.claude.json) or any MCP client:

{

"mcpServers": { "memory": { "command": "/path/to/mcp-memory-rs", "env": { "MCP_MEMORY_CONFIG": "/home/me/.config/mcp-memory-rs/config.toml", "MCP_MEMORY_REQUIRE_CONFIG": "1", "MCP_MEMORY_MODE": "offline", "MCP_DEVICE": "my-laptop" } } } }

Codex (~/.codex/config.toml):

[mcp_servers.memory]

command = "/path/to/mcp-memory-rs" env = { MCP_MEMORY_CONFIG = "/home/me/.config/mcp-memory-rs/config.toml", MCP_MEMORY_REQUIRE_CONFIG = "1", MCP_MEMORY_MODE = "offline", MCP_DEVICE = "my-laptop" }

auto-approve read-only tools (memory_read/list/search/…); writes still gated

default_tools_approval_mode = "approve"

Configuration is TOML — copy config/example.toml to ~/.config/mcp-memory-rs/config.toml and adjust it. The server auto-discovers that XDG-style path. Environment variables override file values (MCP_DEVICE, MCP_MEMORY_MODE, MCP_MEMORY_DIR, …); managed deployments should set MCP_MEMORY_REQUIRE_CONFIG=1 so a missing config fails closed.

Memory model

Memory is organized in categories: named JSON documents (base, projects, my-laptop, workflow_my-laptop, …). Reads return the document plus metadata (content hash, size, last writer, timestamp). Writes replace the category or, with merge=true, patch it per top-level key.

Access control

Multi-device fleets get a small, explicit ACL ([acl] in the config):

Unknown categories are denied for non-admin writers — fail closed.

Fleet sync

Each node runs local-first; one node (or any number) exposes the HTTP API as a sync remote. sync_manifest builds a merkle-rooted fingerprint of all categories; sync_diff compares manifests; memory_sync pushes dirty categories or pulls remote changes. Conflicts between nodes are resolved by the configured conflict_strategy (last-write-wins by default); expected_hash preconditions protect direct writes (MCP and HTTP), not the sync envelope.

The HTTP API is an admin/sync plane: the bearer token grants full access to every category. Per-category ACL applies on the MCP stdio surface, where the device identity is known locally; HTTP callers are identified only by the shared token, so no per-device ACL is enforced there.

The HTTP server requires MCP_MEMORY_TOKEN and refuses to start without it. Bind it to loopback (the default) and tunnel between nodes; do not expose it to the public internet.

MCP_MEMORY_TOKEN=<secret> ./target/release/mcp-memory-rs --http

Tool surface

The same surface is available over HTTP (/api/v1/*) for non-MCP consumers; /health is unauthenticated, everything else requires the bearer token.

AI client compatibility

The server is built to be self-explanatory to a weak client model:

• At initialize it returns an instructions string describing the memory model and the entry-point tools. Some lightweight clients ignore this field; if your client never surfaces it, read the tool descriptions instead — they carry the same guidance (e.g. memory_read takes category, not key).
• Read-only tools are annotated readOnlyHint, so a client such as Codex can auto-approve them. On Codex, an unsupported call / user cancelled result usually means the tool-approval gate fired, not a server fault — set default_tools_approval_mode = "approve" (see the Codex snippet above).
• The server speaks the standard MCP handshake. A client must complete initialize and send notifications/initialized like any MCP client.

Storage layout

With no config and strict mode disabled, standalone base_dir remains ~/.memory for compatibility. The auto-discovered example config uses ~/.local/state/mcp-memory-rs (XDG-style). Config resolution is: explicit MCP_MEMORY_CONFIG, $XDG_CONFIG_HOME, $HOME/.config, legacy cwd-local memory-config.toml, then standalone defaults. Either way the layout is:

<base_dir>/

├── categories/ # one .json file per category — the source of truth ├── backups/ # automatic pre-delete/pre-overwrite backups ├── memory.db # SQLite: FTS5 index, versions, sync state (rebuildable) └── memory.db-wal # write-ahead log (WAL mode; do not copy or delete by hand)

Concurrency (many processes, one store)

Every writer — one stdio server per client plus the optional --http instance — opens its own SQLite connection on the same file. Since 0.3.2 each mutation runs inside a BEGIN IMMEDIATE transaction: SQLite is the mutex that serializes writers *across processes*, so an append (read → modify → write) can never lose another process's entry. Lock acquisition is retried with jittered backoff inside one total budget; when it expires the call fails with a clear Store busy … retry later error instead of hanging or corrupting anything.

[storage]

journal_mode = "wal" # "wal" (default) or "delete" synchronous = "full" # "full" (default, durable) or "normal" sqlite_busy_timeout_ms = 200 # per-statement SQLite busy handler lock_deadline_ms = 10000 # total wait for the write lock per mutation

Category JSON files are written crash-safely: the new content is staged next to the file (a name unique to the transaction), the SQLite rows are committed, then the staged file is renamed into place. A process that dies in between leaves a staged file whose fate is decided by its hash under the next write lock (matching the committed hash → promoted; otherwise → discarded). Readers never rename or delete anything: they serve whichever file carries the committed hash. memory_write merge=true, the log/memory kind check and the optional expected_hash compare-and-swap all run inside the same lock.

Snapshot and restore

memory.db alone is not a backup (WAL, JSON files). Use the CLI:

mcp-memory-rs --snapshot /path/to/snap-2026-09-15 --with-backups  # db + json + backups + manifest

mcp-memory-rs --restore /path/to/snap-2026-09-15 --stopped mcp-memory-rs --verify-store # read-only: db vs json, pending files

--verify-store never opens a Store (no PRAGMA, schema or reconcile): it is safe on a store another binary still owns. A store written by a version older than 0.3.2 can hold categor