Citadel

Historical memory results and configurations

Quick Start

For semantic memory through MCP, install uv and pull the local embedder and optional cross-encoder reranker:

uvx citadeldb-mcp pull e5-large
uvx citadeldb-mcp pull ms-marco-minilm

Set CITADEL_KEY to your vault passphrase (export CITADEL_KEY="your-passphrase" on macOS/Linux or $env:CITADEL_KEY = "your-passphrase" in PowerShell), then start:

uvx citadeldb-mcp --db memory.cdl --embedder e5-large --reranker ms-marco-minilm

The server communicates over stdio. See MCP for client configuration. Model downloads do not need a vault key; serving does.

Memory (Python)

Install the published package with pip install citadeldb. See the Python source-build and semantic-memory guide. Embedders implement embed_with_cancel(texts, cancel_token) and check cancellation between bounded batches. Local Candle models require the candle-embed build feature.

Memory (Rust)

Uses citadeldb and citadeldb-mem with the candle-embed feature. This example loads e5-large and a local cross-encoder reranker. Other presets or a custom Embedder are supported.

use std::sync::Arc;
use citadel::DatabaseBuilder;
use citadel_mem::{AtomInput, CandleEmbedder, CrossEncoder, MemoryEngine, RecallQuery, RerankStrategy};

// Encrypted store (per-atom keys enable cryptographic forgetting)
let db = DatabaseBuilder::new("memory.db")
    .passphrase(b"secret")
    .enable_region_keys(true)
    .create()?;
let mem = MemoryEngine::open(Arc::new(db))?;

let embedder = Arc::new(CandleEmbedder::e5_large("/path/to/e5-large")?);
mem.create_encrypted_region("chat", embedder)?;
mem.set_reranker(
    Arc::new(CrossEncoder::ms_marco_minilm_l6("/path/to/ms-marco-minilm")?),
    RerankStrategy::default(),
);

// Remember raw turns (no LLM)
mem.remember("chat", AtomInput::new("fact", "Alice's cat is named Mochi"))?;
let berlin = mem.remember("chat", AtomInput::new("fact", "Alice lives in Berlin"))?;

// Recall by relevance
for hit in mem.recall("chat", RecallQuery::by_text("where does Alice live?", 5))? {
    println!("{:.3}  {}", hit.relevance.expect("ranked recall"), hit.text);
}

// Cryptographic forgetting: destroy the atom's key
mem.forget_atom("chat", berlin)?;

SQL and key-value

Uses the citadeldb and citadeldb-sql crates - or try SQL with no install in the live playground.

use citadel::DatabaseBuilder;
use citadel_sql::Connection;

let db = DatabaseBuilder::new("my.db")
    .passphrase(b"secret")
    .create()?;

let conn = Connection::open(&db)?;
conn.execute("CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT NOT NULL);")?;
conn.execute("INSERT INTO users (id, name) VALUES (1, 'Alice');")?;
let result = conn.query("SELECT * FROM users;")?;

// Key-value API
let mut wtx = db.begin_write()?;
wtx.insert(b"key", b"value")?;
wtx.commit()?;

let mut rtx = db.begin_read();
assert_eq!(rtx.get(b"key")?.unwrap(), b"value");

// Named tables
let mut wtx = db.begin_write()?;
wtx.create_table(b"sessions")?;
wtx.table_insert(b"sessions", b"token-abc", b"user-42")?;
wtx.commit()?;

// In-memory (no file I/O - useful for testing and WASM)
let mem_db = DatabaseBuilder::new("")
    .passphrase(b"secret")
    .create_in_memory()?;

CLI

citadel --create my.db

citadel> CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT NOT NULL);
citadel> INSERT INTO users (id, name) VALUES (1, 'Alice'), (2, 'Bob');
citadel> SELECT * FROM users;
+----+-------+
| id | name  |
+----+-------+
|  1 | Alice |
|  2 | Bob   |
+----+-------+

citadel> .backup mydb.bak
citadel> .verify
citadel> .upgrade
citadel> .stats
citadel> .audit verify
citadel> .rekey
citadel> .compact clean.db
citadel> .dump users

# P2P sync
citadel> .keygen
citadel> .listen 4248 <KEY>              # Terminal A
citadel> .sync 127.0.0.1:4248 <KEY>      # Terminal B

Citadel Studio

A native desktop client for Windows, macOS, and Linux. Open encrypted vaults, browse tables and memory, run SQL with EXPLAIN and ANALYZE, and inspect vectors and integrity results.

See the Studio guide for screenshots and build instructions. Download Citadel Studio for Windows, macOS, or Linux.

Agent frameworks

The adapters implement framework-specific storage, session, and retrieval interfaces. Each requires an explicit embedder. See the package README for setup, search behavior, and supported filters.

pip install citadeldb-langgraph

One database serves every adapter on the thread that opened it, so a graph's long-term store and its session transcripts can share one encrypted file. See packaging/ for each package's own README.

MCP

Serve an encrypted memory region to Claude Desktop or any MCP client. citadeldb-mcp is published to PyPI and listed in the official MCP registry as dev.citadeldb/mcp. Run it without installing through uvx.

For the recommended semantic-recall setup, pull the embedder and cross-encoder reranker once:

uvx citadeldb-mcp pull e5-large
uvx citadeldb-mcp pull ms-marco-minilm

The pull commands do not need a vault key. Before starting the server, set CITADEL_KEY to the vault passphrase: use export CITADEL_KEY="your-passphrase" on macOS/Linux or $env:CITADEL_KEY = "your-passphrase" in PowerShell. Then run:

uvx citadeldb-mcp --db memory.cdl --embedder e5-large --reranker ms-marco-minilm

--db, --embedder, and CITADEL_KEY are required when serving. The reranker is optional, but e5-large with ms-marco-minilm is the configuration used for the memory benchmarks.

To install the executable instead, run pip install citadeldb-mcp or cargo install citadeldb-mcp. Pull the same models with citadeldb-mcp pull e5-large and citadeldb-mcp pull ms-marco-minilm, then add it to claude_desktop_config.json:

{
  "mcpServers": {
    "citadel": {
      "command": "citadeldb-mcp",
      "args": [
        "--db", "/absolute/path/to/memory.cdl",
        "--embedder", "e5-large",
        "--reranker", "ms-marco-minilm"
      ],
      "env": { "CITADEL_KEY": "your-passphrase" }
    }
  }
}

Historical memory benchmarks

Recorded LoCoMo and LongMemEval results are summarized below; their configurations and limitations predate the current memory-engine changes. SQL comparisons with unencrypted SQLite across 59 cases are under Speed benchmarks.

LoCoMo - gpt-4o-mini reader and judge with the harness's prompts, mean of 3 runs measured August 18, 2026:

Retrieval is identical across the three runs; the spread is reader and judge nondeterminism. A manual audit estimates that ~6.4% of LoCoMo answer keys are erroneous, so raw accuracy should be interpreted with that annotation noise in mind.

Memory is built with no LLM - raw turns enriched with supplied photo captions and image-search text, indexed and recalled deterministically.

LongMemEval_S (arXiv 2410.10813) full-haystack split (~40-50 sessions/question), gpt-4o reader, official CoT prompt and gpt-4o-2024-08-06 judge:

Full-haystack stresses retrieval against distractors (not the oracle reader ceiling). Protocol and per-type results in citadel-membench.

Encrypted memory engine

The same encrypted pages that hold SQL tables also hold memory. Three crates make up the memory engine:

• citadeldb-vector - a VECTOR(N) SQL type, distance operators (<-> L2, <#> inner, <=> cosine), and a PRISM-backed filtered ANN index that reads through the encrypted page store.
• citadeldb-mem - the memory engine (regions, atoms, edges) with hybrid recall and cryptographic forgetting: an atom or region is erased by destroying its key, at whole-store, per-region, and per-atom granularity.
• citadeldb-mcp - a Model Context Protocol server exposing a Citadel memory region (encrypted by default) to any MCP client (Claude Desktop, IDEs) as recall/remember/link/evolve/forget/verify tools.

Zero-LLM memory path

citadeldb-mem stores raw conversation content without a summarizer LLM. Recall uses embeddings, BM25 keyword matching, and an optional reranker. Local embedding and reranking backends keep this processing on-device; custom backends determine their own network use and costs. The benchmark readers and judges are separate LLMs - gpt-4o-mini for LoCoMo, gpt-4o for LongMemEval. The protocol and results are in citadel-membench.

Agent runtime

• citadeldb-llm - the provider-neutral LLM client layer (Claude, OpenAI, Ollama, Gemini) behind one factory, with canonical request hashing and a non-secret client request identity.
• citadeldb-ai - an autonomous agent runtime (ReAct + Reflexion, tool registry, budget caps, pluggable LLM backends) that uses citadeldb-mem for persistence.

Features

• Encrypted at rest - AES-256-CTR + HMAC-SHA256 per page, verified before decryption
• SQL - JOINs, subqueries, CTEs (recursive + WITH-DML), UNION/INTERSECT/EXCEPT, window functions, views, materialized views, triggers, TEMP tables, generated columns (STORED + VIRTUAL), constraints, full FK actions, UPSERT, RETURNING, JSON/JSONB (14 Postgres operators + SQL/JSON path language), full-text search, prepared statements with plan caching, and a queryable system catalog. Full list under SQL
• ACID - Copy-on-Write B+ tree, shadow paging, no WAL. Snapshot isolation with concurrent readers
• Authenticated commit slots - the commit metadata (table roots, catalog) carries its own HMAC; older files migrate one-way via .upgrade
• P2P sync - Merkle-based table diffing over Noise-encrypted channels with PSK auth
• CLI - SQL shell with tab completion, syntax highlighting, 27 dot-commands (.backup, .verify, .upgrade, .rekey, .sync, .dump, ...)
• Citadel Studio - Native desktop client for SQL, stored memory, vector inspection, and vault diagnostics
• 3-tier key hierarchy - Passphrase -> Argon2id -> Master Key -> AES-KW -> REK -> HKDF -> DEK + MAC
• Cryptographic forgetting - Whole-store and per-region / per-atom key erasure via citadeldb-mem. Pre-erasure backups, copied keys, and exported plaintext are outside that erasure
• FIPS-oriented at-rest profile - PBKDF2-HMAC-SHA256 + AES-256-CTR for database storage; not a claim of whole-product validation
• Audit log - HMAC-SHA256 chained within files and across retained v2 generations; retained-history verification detects record edits and broken retained links, but there is no external anti-rollback anchor
• Hot backup - Consistent snapshots via MVCC, no write blocking
• Overflow pages - Large values handled transparently, up to 1 GiB per value
• Cross-platform - Windows, Linux, macOS. Python, C FFI, and WebAssembly bindings
• Thousands of tests - Unit, integration, and torture tests across the workspace

Speed benchmarks

Measured on September 13, 20 and 27, 2026 (UTC) on an Intel Core i9-12900HX, Windows 11 Pro, Rust 1.98.0, and SQLite 3.51.3. Runs use one fixed logical processor, with durability disabled and both caches configured for 4,096 pages (about 32 MiB). Most cases use 100K rows; schemas and operations vary as listed below.

Each time is the arithmetic mean of two or four per-run sample medians, with 30 samples per run. Ratios use unrounded SQLite time / Citadel time: above 1 means Citadel is faster, below 1 means Citadel is slower. For example, 0.5x means Citadel takes twice as long as SQLite.

Sixteen comparisons were refreshed on September 27 at 0dbe126c: nine execution cases and seven cached repeat reads. Each refreshed row uses four new runs per engine in Citadel/SQLite/SQLite/Citadel, then SQLite/Citadel/Citadel/SQLite order. Other rows retain their September 13 or 20 measurements and source revisions. This is a combined snapshot, not a full-suite run at the latest revision. Source revisions, run settings, medians, 95% intervals, and drift identify every row.

Execution speed

37 comparisons of writes and reads that execute each iteration, including rotating-parameter queries. Fixture resets are excluded unless the case description says otherwise.

Benchmark                     Citadel        SQLite         Ratio
----------------------------------------------------------------------
join_param                    2.6 us         51.2 us        19.7x
fts_rank_first_execution      6.86 ms        64.5 ms        9.4x
insert_returning              80.8 us        351 us         4.34x
update_returning              65.7 us        232 us         3.53x
window_agg                    33.6 ms        108 ms         3.2x
upsert_returning              124 us         363 us         2.94x
sort_paginate_pk              9.13 us        26.1 us        2.86x
delete_returning              97.2 us        269 us         2.76x
window_rank                   69.2 ms        182 ms         2.63x
fts_phrase                    5.66 ms        14.6 ms        2.58x
fts_match                     4.94 ms        12.1 ms        2.44x
json_extract                  22.6 ms        49 ms          2.17x
scan                          6.31 ms        13.2 ms        2.09x
insert                        25.5 us        51.8 us        2.03x
insert_gen_virtual            36.4 us        66.4 us        1.82x
wide_proj_full                6.83 ms        12.1 ms        1.77x
insert_gen_stored             37.3 us        65.7 us        1.76x
upsert_all_new                36.2 us        63.5 us        1.76x
wide_proj_pk                  416 us         728 us         1.75x
truncate                      57.6 us        101 us         1.75x
upsert_dedup                  31.9 us        53.7 us        1.68x
savepoint_rollback            2.08 ms        3.22 ms        1.55x
delete                        75.1 us        116 us         1.54x
wide_proj_2col                651 us         998 us         1.53x
covered_count                 377 us         561 us         1.49x
wide_proj_3col                1.28 ms        1.89 ms        1.47x
savepoint_nested              232 us         322 us         1.39x
update                        36.1 us        45.8 us        1.27x
insert_select                 171 us         214 us         1.25x
with_dml                      122 us         147 us         1.21x
fk_cascade_delete_only        52.4 us        63.2 us        1.2x
fk_cascade                    122 us         144 us         1.18x
savepoint_create              916 ns         1.07 us        1.16x
upsert_mixed                  56.1 us        64.7 us        1.15x
covered_range                 105 us         119 us         1.14x
update_gen_propagate          65.7 us        70.8 us        1.08x
upsert_counter                79.1 us        82.7 us        1.05x

Cached repeat reads

22 comparisons of identical reads against unchanged data. Citadel reuses cached results; union reuses projected branch rows and reconstructs UNION ALL output. SQLite executes the query again. These timings do not represent the first query after a write.

Benchmark                     Citadel        SQLite         Ratio
----------------------------------------------------------------------
correlated_in                 242 ns         2.78 s         11500000x
fts_rank                      534 ns         63.8 ms        120000x
correlated_exists             239 ns         9.61 ms        40200x
jsonb_contains                1.81 us        40.5 ms        22400x
sort_nocase                   446 ns         4.71 ms        10600x
cte                           1.46 us        9.29 ms        6350x
sort                          674 ns         4.03 ms        5990x
group_by                      2.47 us        14.5 ms        5860x
sum                           529 ns         2.74 ms        5180x
distinct                      1.82 us        5.94 ms        3260x
full_outer_join               25.4 us        31.7 ms        1250x
correlated_scalar             23.6 us        28.1 ms        1190x
recursive_cte                 267 ns         175 us         654x
partial_index_point           269 ns         22.6 us        84.1x
view_point                    300 ns         22.8 us        75.9x
point                         302 ns         22.6 us        74.9x
filter                        38.6 us        2.74 ms        70.9x
view_filter                   38.6 us        2.65 ms        68.8x
count                         855 ns         37.4 us        43.7x
select_gen_virtual            2.21 us        34.5 us        15.6x
join                          24.7 us        147 us         5.94x
union                         50.6 us        230 us         4.54x

Citadel-only

No SQLite comparison is reported for these seven cases. json_table executes each iteration; the other six measure cached repeat reads.

Benchmark                     Citadel        SQLite         Ratio
----------------------------------------------------------------------
json_table                    7.42 ms        -              -
lateral                       2.63 us        -              -
date_sort                     1.81 us        -              -
date_extract                  848 ns         -              -
date_groupby                  576 ns         -              -
date_arith                    260 ns         -              -
date_range_scan               257 ns         -              -

Index comparisons

The same query within Citadel, with and without its index. Ratios are unindexed / indexed time. json_gin rotates unique JSON-id probes; fts_index repeats a fixed query on a TEXT column. Both execute each iteration.

Benchmark                     Without index  With index     Ratio
----------------------------------------------------------------------
json_gin                      8.26 ms        5.77 us        1430x
fts_index                     1.95 s         4.76 ms        409x

cargo bench --locked -p citadeldb-sql --bench h2h_bench -- \
  '^(update|update_gen_propagate|update_returning)/citadel/$' \
  --sample-size 30 --warm-up-time 3 --measurement-time 8 --noplot

SQL

Statements - CREATE/DROP TABLE (incl. TEMP), ALTER TABLE (ADD/DROP/RENAME COLUMN, RENAME TABLE, DISABLE/ENABLE TRIGGER), CREATE/DROP INDEX (incl. partial WHERE, expression keys, CONCURRENTLY), REINDEX [DATABASE] / REINDEX [TABLE | INDEX] name, CREATE/DROP VIEW, CREATE/DROP MATERIALIZED VIEW (with REFRESH [CONCURRENTLY]), CREATE/DROP TRIGGER (BEFORE/AFTER/INSTEAD OF, FOR EACH ROW/STATEMENT, REFERENCING NEW/OLD TABLE, WHEN, UPDATE OF cols), INSERT (VALUES, SELECT, ON CONFLICT DO NOTHING/DO UPDATE, ON CONSTRAINT), SELECT, UPDATE (including correlated subqueries in SET), DELETE, TRUNCATE TABLE, RETURNING (with OLD/NEW), BEGIN [READ ONLY | READ WRITE]/COMMIT/ROLLBACK, SAVEPOINT/RELEASE/ROLLBACK TO, SET [LOCAL] TIME ZONE, EXPLAIN, REFRESH MATERIALIZED VIEW

Constraints - PRIMARY KEY, NOT NULL, UNIQUE, DEFAULT, CHECK (column + table level), FOREIGN KEY with full referential actions (ON DELETE / ON UPDATE CASCADE / SET NULL / SET DEFAULT / RESTRICT / NO ACTION), GENERATED ALWAYS AS (...) STORED|VIRTUAL

Collations - BINARY, NOCASE (ASCII case-insensitive), and RTRIM (ignores trailing spaces). Text primary keys use their declared collation; foreign keys use the referenced columns' collation. Column index keys inherit their column's collation unless overridden with COLLATE. INTERVAL keys (primary, unique, foreign and index) compare by length as = does, so '1 month' and '30 days' are one key; REINDEX converts interval columns stored by earlier versions.

Types - INTEGER, REAL, TEXT, BLOB, BOOLEAN, DATE, TIME, TIMESTAMP (WITH TIME ZONE), INTERVAL, JSON, JSONB, TSVECTOR, TSQUERY, ARRAY

JSON / JSONB - Postgres operators plus SQL/JSON path functions and the SQL:2023 item methods .bigint(), .decimal(), .integer(), .number(), .string(), .boolean(), .date(), .time(), .time_tz(), .timestamp(), and .timestamp_tz(). Time-zone-dependent evaluation uses the connection's transactional SET [LOCAL] TIME ZONE context.

Clauses - JOINs (INNER, LEFT, RIGHT, CROSS, FULL OUTER; LATERAL with CROSS/INNER/LEFT), subqueries (scalar, IN, EXISTS, correlated), CTEs (WITH / WITH RECURSIVE / WITH-DML: WITH x AS (INSERT/UPDATE/DELETE ... [RETURNING *]) SELECT ...), UNION/INTERSECT/EXCEPT [ALL], CASE, BETWEEN, LIKE, DISTINCT, ANY / ALL (subquery + array forms), GROUP BY/HAVING, ORDER BY, LIMIT/OFFSET

Window functions - ROW_NUMBER, RANK, DENSE_RANK, NTILE, LAG, LEAD, FIRST_VALUE, LAST_VALUE, SUM/COUNT/AVG/MIN/MAX OVER with PARTITION BY, ORDER BY, ROWS/RANGE frames. In grouped queries, window functions operate on the groups remaining after GROUP BY and HAVING.

Views - CREATE/DROP VIEW, OR REPLACE, IF NOT EXISTS/IF EXISTS, column aliases, nested views

Materialized views - CREATE MATERIALIZED VIEW [IF NOT EXISTS] name AS SELECT ..., REFRESH MATERIALIZED VIEW [CONCURRENTLY] name (CONCURRENTLY does a diff-merge - DELETE removed rows, UPDATE changed rows, INSERT new rows - instead of TRUNCATE+repopulate), DROP MATERIALIZED VIEW [CASCADE], full backing-table semantics (indexes, joins, planner sees a real table), pg_matviews introspection

Triggers - CREATE TRIGGER name {BEFORE|AFTER|INSTEAD OF} {INSERT|UPDATE [OF cols]|DELETE} ON table FOR EACH {ROW|STATEMENT} [REFERENCING NEW TABLE AS new_t OLD TABLE AS old_t] [WHEN (expr)] BEGIN ... END. INSTEAD OF triggers make views writable. Transition tables work as virtual tables in trigger bodies. ALTER TABLE ... DISABLE/ENABLE TRIGGER [name|ALL]. PG-faithful name-order firing. Introspection via information_schema.triggers and SHOW TRIGGERS [ON table].

TEMP tables - CREATE TEMP TABLE ... lives in a per-connection in-memory database, dropped on disconnect. Full DDL/DML/index/constraint/trigger parity with persistent tables.

Functions - COUNT, SUM, AVG, MIN, MAX, LENGTH, UPPER, LOWER, SUBSTR/SUBSTRING, TRIM/LTRIM/RTRIM, REPLACE, INSTR, CONCAT, HEX, ABS, ROUND, CEIL/CEILING, FLOOR, SIGN, SQRT, RANDOM, COALESCE, NULLIF, GREATEST, LEAST, CAST, TYPEOF, IIF. Non-window aggregates support FILTER (WHERE ...).

Date/Time Functions - NOW, CURRENT_TIMESTAMP, CURRENT_DATE, CURRENT_TIME, LOCALTIMESTAMP, LOCALTIME, CLOCK_TIMESTAMP, EXTRACT, DATE_PART, DATE_TRUNC, DATE_BIN, AGE, MAKE_DATE, MAKE_TIME, MAKE_TIMESTAMP, MAKE_INTERVAL, JUSTIFY_DAYS, JUSTIFY_HOURS, JUSTIFY_INTERVAL, ISFINITE, DATE, TIME, DATETIME, STRFTIME, JULIANDAY, UNIXEPOCH, TIMEDIFF, AT TIME ZONE. Supports INTERVAL '1 year 2 months', DATE '2024-01-15', TIMESTAMP '2024-01-15 12:30:00Z', infinity/-infinity sentinels, BC dates, full IANA zone parsing (jiff), PG-normalized INTERVAL comparison.

Full-text search - tsvector / tsquery types, to_tsvector / to_tsquery / plainto_tsquery / phraseto_tsquery / websearch_to_tsquery builders, @@ match operator, ts_rank / ts_rank_cd ranking with weighted positions (A/B/C/D), prefix matching (term:*), phrase distance (<N>), inverted indexes via CREATE INDEX ... USING fts

System catalog - information_schema.tables, information_schema.columns, information_schema.key_column_usage, information_schema.table_constraints, information_schema.triggers, pg_timezone_names, pg_timezone_abbrevs, pg_matviews (virtual tables, queryable). SHOW TRIGGERS [ON table] and SHOW MATERIALIZED VIEWS shorthands for the corresponding catalog queries.

Prepared statements - $1, $2, ... positional parameters with LRU statement cache plus snapshot-tagged plan caching for joins and compound queries (cache invalidates only on commit, never per-call)

Multi-statement scripts - Connection::execute_script(sql) runs ;-separated statements in one call, returning per-statement outcomes with partial-success preserved. WASM: db.run(sql) returns [{type, ...}, ...].

UPSERT - INSERT ... ON CONFLICT (cols) DO NOTHING / DO UPDATE SET col = excluded.col ... WHERE ... and ON CONFLICT ON CONSTRAINT idx_name. excluded.* refers to the proposed row; bare col refers to the existing row.

Security

No plaintext on disk. Every page is encrypted before writing and authenticated before reading.

Separate key file. Encryption keys live in {dbname}.citadel-keys, not inside the database. The passphrase derives a master key in memory via Argon2id (or PBKDF2 in the FIPS-oriented at-rest profile) and never touches disk.

Key backup. Export an encrypted key backup with a separate recovery passphrase. Restore access without re-encrypting the entire database.

Instant rekey. Changing the passphrase re-wraps the root encryption key. No page re-encryption - instant regardless of database size.

Encrypted sync. Noise protocol (NNpsk0_25519_ChaChaPoly_BLAKE2s) with a 256-bit pre-shared key. Ephemeral Curve25519 keys per session for forward secrecy.

Architecture

Clients and bindings:
+---------------------------------------------+
|               citadel-studio                |  Memory, SQL, and vault client
+----------------------+----------------------+
|     citadel-cli      |    citadel-python    |  CLI, Python wheel
+----------------------+----------------------+
|     citadel-ffi      |     citadel-wasm     |  C FFI, WebAssembly
+----------------------+----------------------+

Agent layer:
+---------------------------------------------+
|                 citadel-ai                  |  Agent runtime (ReAct + Reflexion)
+---------------------------------------------+
|                 citadel-llm                 |  LLM clients: Claude, OpenAI, Ollama, Gemini
+---------------------------------------------+

Memory layer:
+---------------------------------------------+
|                 citadel-mcp                 |  MCP server for memory tools
+---------------------------------------------+
|                 citadel-mem                 |  Regions, atoms, recall, erasure
+---------------------------------------------+
|                citadel-vector               |  VECTOR(N) type + PRISM filtered ANN
+---------------------------------------------+

Encrypted database engine:
+----------------------+----------------------+
|     citadel-sql      |    sql-json-path     |  SQL frontend, SQL/JSON paths
+----------------------+----------------------+
|                   citadel                   |  Database API, builder, vault lifecycle
+-------------+--------------+----------------+
| citadel-txn | citadel-sync | citadel-crypto |  Transactions, replication, keys
+-------------+--------------+----------------+
|       citadel-buffer       |  citadel-page  |  Buffer pool (SIEVE), page codec
+----------------------------+----------------+
|                 citadel-io                  |  File I/O, fsync, io_uring
+---------------------------------------------+
|                citadel-core                 |  Types, errors, cancellation
+---------------------------------------------+

Evaluation harnesses:
+----------------------+----------------------+
|   citadel-membench   |     citadel-swe      |  Memory and agent benchmarks
+----------------------+----------------------+

Studio calls the database and SQL APIs directly and uses MemoryMaintenance for stored-memory inspection and erasure. It needs no MCP server or embedding model.

Page Layout (8,208 bytes)

+----------+--------------------+----------+
|  IV 16B  |  Ciphertext 8160B  |  MAC 32B |
+----------+--------------------+----------+

Fresh random IV per page. HMAC verified before decryption.

Commit Protocol

Shadow paging with a god byte - one byte selects the active commit slot. Atomic commits without WAL:

1. Write dirty pages to new locations (CoW)
2. Compute Merkle hashes bottom-up
3. Update the inactive commit slot
4. Flip the god byte

SyncMode::Full flushes the pages and commit slot before the flip, then flushes the selector before returning success. If that final flush fails, the commit's durability is uncertain and further writes return Error::ReopenRequired. Close and reopen the database before writing again.

Integrity Boundary

What the at-rest integrity machinery does and does not guarantee against an attacker with file access:

• Per-page HMAC binds (epoch, page_id, IV, ciphertext). Any modification of a page's bytes is detected before decryption. It does not bind the commit generation: a page image validly written in the past for the same (page_id, epoch) verifies forever.
• Commit slots have two accepted formats. V1 slots carry a truncated HMAC-SHA256 over every field except the MAC itself; legacy slots carry only a keyless checksum over a prefix. Checksum-valid legacy slots remain readable only while no V1 requirement is recorded. Once both physical slots are valid V1 and the vault records that one-way requirement, any checksum-valid legacy slot is rejected as downgrade evidence, and writers refuse to create one.
• Rollback to an older genuine state is outside this boundary. An earlier authenticated slot plus its matching pages can pass the data-file checks; an older internally consistent snapshot of all local vault state, including the data, key, and retained audit files, also passes local authentication. Detecting freshness requires an external anchor - for example, store the latest commit's txn_id and Merkle root outside the attacker's reach and compare them after opening.

Language Bindings

C / C++

Static or dynamic library with auto-generated citadel.h (cbindgen). Exported entry points are panic-safe.

#include "citadel.h"

int main(void) {
    struct CitadelDb *db = NULL;
    struct CitadelSqlConn *conn = NULL;
    struct CitadelSqlResult *result = NULL;
    citadel_error_t status = citadel_create(
        "my.db", (const uint8_t *)"secret", 6, NULL, &db);
    if (status != CITADEL_ERROR_T_OK) goto cleanup;

    status = citadel_sql_open(db, &conn);
    if (status != CITADEL_ERROR_T_OK) goto cleanup;
    status = citadel_sql_execute(conn, "SELECT 1 + 1 AS value;", &result);

cleanup:
    citadel_sql_result_free(result);
    citadel_sql_close(conn);
    citadel_close(db);
    return status == CITADEL_ERROR_T_OK ? 0 : 1;
}

WebAssembly

Install with npm install @citadeldb/wasm.

import init, { CitadelDb } from "@citadeldb/wasm";

await init();

const db = new CitadelDb("secret");
db.execute("CREATE TABLE t (id INTEGER PRIMARY KEY, name TEXT);");
db.execute("INSERT INTO t (id, name) VALUES (1, 'Alice');");

const result = db.query("SELECT * FROM t;");
// { columns: ["id", "name"], rows: [[1, "Alice"]] }

db.put(new Uint8Array([1, 2, 3]), new Uint8Array([4, 5, 6]));
db.free();

Build the npm package: bash scripts/publish-wasm.sh

Python

One importable wheel with the full engine (SQL, vectors, memory, agent runtime) and bundled type stubs.

pip install citadeldb

import citadeldb

db = citadeldb.connect("my.db", key="secret", create=True)
db.execute("CREATE TABLE t (id INTEGER PRIMARY KEY, name TEXT)")
db.execute("INSERT INTO t VALUES (1, 'Alice')")
db.query("SELECT * FROM t").to_dicts()
# [{'id': 1, 'name': 'Alice'}]

Building

Rust 1.95+.

git clone https://github.com/yp3y5akh0v/citadel.git
cd citadel
cargo build --release

Feature Flags

License

Apache-2.0