mcp_memory/
lib.rs

1pub mod actions;
2#[cfg(feature = "code")]
3pub mod code;
4#[cfg(feature = "code")]
5pub mod code_registry;
6pub mod config;
7pub mod errors;
8pub mod http;
9pub mod ivf;
10pub mod kg;
11pub mod protocol;
12pub mod server;
13pub mod tls;
14pub mod tools;
15pub mod types;
16pub mod watcher;
17pub mod vector_actions;
18pub mod vector_store;
19
20use clap::{Parser, ValueEnum};
21use usearch::{MetricKind, ScalarKind};
22use vector_store::{IndexKind, VectorConfig};
23
24/// ANN index backend selectable from the CLI.
25#[derive(Clone, Copy, Debug, ValueEnum)]
26pub enum VecIndex {
27    /// usearch HNSW graph index (default): best recall/latency.
28    Hnsw,
29    /// IVF-Flat: k-means partitioned, lighter memory, fast to build/rebuild.
30    Ivf,
31}
32
33impl From<VecIndex> for IndexKind {
34    fn from(v: VecIndex) -> Self {
35        match v {
36            VecIndex::Hnsw => IndexKind::Hnsw,
37            VecIndex::Ivf => IndexKind::Ivf,
38        }
39    }
40}
41
42/// Distance metric for the vector index.
43#[derive(Clone, Copy, Debug, ValueEnum)]
44pub enum VecMetric {
45    /// Cosine similarity (default; good for normalized embeddings).
46    Cos,
47    /// Inner / dot product.
48    Ip,
49    /// Squared Euclidean (L2) distance.
50    L2sq,
51}
52
53impl From<VecMetric> for MetricKind {
54    fn from(m: VecMetric) -> Self {
55        match m {
56            VecMetric::Cos => MetricKind::Cos,
57            VecMetric::Ip => MetricKind::IP,
58            VecMetric::L2sq => MetricKind::L2sq,
59        }
60    }
61}
62
63/// Scalar representation stored in the index (lower precision = less memory).
64#[derive(Clone, Copy, Debug, ValueEnum)]
65pub enum VecQuant {
66    /// 32-bit float (default; full precision).
67    F32,
68    /// 16-bit half-precision IEEE float (half the memory, slight recall loss).
69    F16,
70    /// 16-bit brain float (half the memory, wider range than f16).
71    Bf16,
72    /// 8-bit integer quantization (quarter the memory).
73    I8,
74}
75
76impl From<VecQuant> for ScalarKind {
77    fn from(q: VecQuant) -> Self {
78        match q {
79            VecQuant::F32 => ScalarKind::F32,
80            VecQuant::F16 => ScalarKind::F16,
81            VecQuant::Bf16 => ScalarKind::BF16,
82            VecQuant::I8 => ScalarKind::I8,
83        }
84    }
85}
86
87/// Wire transport the server listens on. The JSON-RPC/MCP semantics are
88/// identical across all three — only the framing differs.
89#[derive(Copy, Clone, Debug, PartialEq, Eq, ValueEnum)]
90#[clap(rename_all = "lowercase")]
91pub enum Transport {
92    /// Newline-delimited JSON-RPC over stdin/stdout (default; for Claude
93    /// Desktop / Claude Code and other process-spawning clients).
94    Stdio,
95    /// Newline-delimited JSON-RPC over a TCP socket (one message per line),
96    /// accepting many concurrent connections.
97    Tcp,
98    /// MCP Streamable HTTP: POST JSON-RPC to `/mcp` (responses as JSON or, when
99    /// the client `Accept`s it, an SSE stream), plus `GET /mcp` for a standalone
100    /// server→client SSE stream.
101    Http,
102}
103
104#[derive(Parser, Debug)]
105#[command(name = "MCP Memory Server")]
106#[command(about = "Knowledge graph memory server for MCP — entities, relations, and observations persisted in SQLite with FTS5 search", long_about = None)]
107pub struct Args {
108    /// Path to the memory file
109    #[arg(short = 'f', long = "memory-file")]
110    pub memory_file: Option<String>,
111
112    /// Transport to listen on: stdio, tcp, or http
113    #[arg(short = 't', long = "transport", value_enum, default_value_t = Transport::Stdio)]
114    pub transport: Transport,
115
116    /// Address to bind for the `tcp` and `http` transports
117    #[arg(short = 'b', long = "bind", default_value = "127.0.0.1:8080")]
118    pub bind: String,
119
120    /// Log level
121    #[arg(short, long, default_value = "info")]
122    pub log_level: String,
123
124    /// Bearer token required on the `tcp` (first line) and `http`
125    /// (`Authorization` header) transports. Overrides `--auth-token-file` and
126    /// the `MCP_MEMORY_AUTH_TOKEN` env var. stdio is never authenticated.
127    #[arg(long = "auth-token")]
128    pub auth_token: Option<String>,
129
130    /// Path to a file whose trimmed contents are the bearer token. An empty
131    /// file is rejected (fail closed). Ignored if `--auth-token` is set.
132    #[arg(long = "auth-token-file")]
133    pub auth_token_file: Option<String>,
134
135    /// SQLite mmap size in bytes (default: 64 MiB).
136    #[arg(long = "mmap-size", default_value_t = 67108864)]
137    pub mmap_size: i64,
138
139    /// SQLite page size in bytes; power of two (default: 4096, matches the Linux
140    /// page / filesystem block size). Only applies to a freshly-created database.
141    #[arg(long = "page-size", default_value_t = 4096)]
142    pub page_size: i64,
143
144    /// SQLite page cache size in MiB (default: 32).
145    #[arg(long = "cache-size-mb", default_value_t = 32)]
146    pub cache_size_mb: i64,
147
148    /// SQLite busy timeout in milliseconds (default: 5000).
149    #[arg(long = "busy-timeout-ms", default_value_t = 5000)]
150    pub busy_timeout_ms: u64,
151
152    /// Interval in milliseconds for a background `wal_checkpoint(PASSIVE)` that
153    /// bounds the durability window in async mode (default: 500). 0 disables it.
154    #[arg(long = "wal-flush-ms", default_value_t = 500)]
155    pub wal_flush_ms: u64,
156
157    /// Entity-metadata LRU cache capacity (0 falls back to 10000).
158    #[arg(long = "lru-cache-size", default_value_t = 10000)]
159    pub lru_cache_size: usize,
160
161    /// Number of read-only SQLite connections backing concurrent reads. WAL
162    /// mode allows readers to run in parallel with each other and the single
163    /// writer; a larger pool raises read concurrency at the cost of a little
164    /// memory (each connection carries its own page cache). `0` (default)
165    /// auto-scales to the CPU count, clamped to [1, 32].
166    #[arg(long = "read-pool-size", default_value_t = 4)]
167    pub read_pool_size: usize,
168
169    /// Path to a PEM certificate chain to serve the `http` transport over TLS
170    /// (HTTPS). Requires --tls-key. Falls back to the MCP_TLS_CERT env var.
171    /// When unset, the `http` transport stays plaintext.
172    #[arg(long = "tls-cert")]
173    pub tls_cert: Option<String>,
174
175    /// Path to the PEM private key matching --tls-cert. Falls back to the
176    /// MCP_TLS_KEY env var.
177    #[arg(long = "tls-key")]
178    pub tls_key: Option<String>,
179
180    /// Enable vector / semantic search: exposes the `vector_*` and
181    /// `hybrid_search` tools backed by a usearch HNSW index. Off by default
182    /// (a pure knowledge-graph server). The `--embedding-dims` / `--vec-*` flags
183    /// only take effect when this is set.
184    #[arg(long = "vectors", default_value_t = false)]
185    pub vectors: bool,
186
187    /// Enable tree-sitter code-symbol indexing: exposes the `code_index`,
188    /// `code_outline`, `code_search`, and `code_get_symbol` tools, which parse
189    /// source files and store functions/classes/methods (and their
190    /// call/define edges) in the graph. On by default (when built with the `code`
191    /// feature, which is on by default).
192    #[arg(long = "code", default_value_t = true)]
193    pub code: bool,
194
195    /// Embedding dimension for vector search (default: 384). Requires --vectors.
196    #[arg(long = "embedding-dims", default_value_t = 384)]
197    pub embedding_dims: u32,
198
199    /// Distance metric for the vector index. Requires --vectors.
200    #[arg(long = "vec-metric", value_enum, default_value_t = VecMetric::Cos)]
201    pub vec_metric: VecMetric,
202
203    /// Scalar quantization for the vector index (lower = less memory). Requires --vectors.
204    #[arg(long = "vec-quantization", value_enum, default_value_t = VecQuant::F32)]
205    pub vec_quantization: VecQuant,
206
207    /// HNSW graph degree `M` (higher = better recall, more memory). Requires --vectors.
208    #[arg(long = "vec-connectivity", default_value_t = 16)]
209    pub vec_connectivity: usize,
210
211    /// HNSW `efConstruction` (higher = better index quality, slower inserts). Requires --vectors.
212    #[arg(long = "vec-expansion-add", default_value_t = 200)]
213    pub vec_expansion_add: usize,
214
215    /// HNSW `efSearch` (higher = better recall, slower queries). Requires --vectors.
216    #[arg(long = "vec-expansion-search", default_value_t = 50)]
217    pub vec_expansion_search: usize,
218
219    /// ANN index backend: `hnsw` (default) or `ivf` (IVF-Flat). Requires --vectors.
220    #[arg(long = "vec-index", value_enum, default_value_t = VecIndex::Hnsw)]
221    pub vec_index: VecIndex,
222
223    /// IVF: number of Voronoi cells / centroids (default: 256). Requires --vec-index ivf.
224    #[arg(long = "ivf-nlist", default_value_t = 256)]
225    pub ivf_nlist: usize,
226
227    /// IVF: cells probed per query — higher = better recall, slower (default: 8).
228    /// Requires --vec-index ivf.
229    #[arg(long = "ivf-nprobe", default_value_t = 8)]
230    pub ivf_nprobe: usize,
231}
232
233impl Args {
234    /// Build the vector index configuration from the `--embedding-dims` /
235    /// `--vec-*` / `--ivf-*` flags. Only meaningful when `--vectors` is set.
236    pub fn vector_config(&self) -> VectorConfig {
237        VectorConfig {
238            dims: self.embedding_dims,
239            index_kind: self.vec_index.into(),
240            metric: self.vec_metric.into(),
241            quantization: self.vec_quantization.into(),
242            connectivity: self.vec_connectivity,
243            expansion_add: self.vec_expansion_add,
244            expansion_search: self.vec_expansion_search,
245            ivf_nlist: self.ivf_nlist,
246            ivf_nprobe: self.ivf_nprobe,
247        }
248    }
249}
mcp_memory/lib.rs

mcp_memory/
lib.rs
