Architecture

System Overview

┌─────────────────────────────────────────────────────┐
│                    CLI (clap)                        │
│  uteke-cli crate — auto-routes to server if running │
├─────────────────────────────────────────────────────┤
│              HTTP API (uteke-serve)                  │
│  /health /remember /recall /search /list /forget    │
│  /stats /namespaces /room/* /mcp — CORS, ~42ms     │
├─────────────────────────────────────────────────────┤
│                    Uteke API                         │
│          uteke-core crate (lib)                      │
├──────────┬──────────┬───────────┬────────────────────┤
│   ONNX   │  usearch │   FTS5    │      SQLite        │
│ Embedding│ Vector   │ Full-Text │   Metadata Store   │
│ (768d)   │ Index    │  Search   │    (rusqlite)      │
│          │ (HNSW)   │ (virtual  │                    │
│          │ RwLock   │   table)  │                    │
├──────────┴──────────┴───────────┴────────────────────┤
│              ~/.codecora/uteke/ (local storage)               │
│ uteke.db │ uteke_index.usearch │ embeddinggemma-q4/ │
└─────────────────────────────────────────────────────┘

Workspace Crates

Crate	Purpose
uteke-core	Library — storage, embedding, vector search, FTS5, operations
uteke-cli	CLI binary — clap commands, JSON output, server proxy
uteke-server	HTTP server — persistent daemon for fast agent access
uteke-mcp	MCP JSON-RPC server — stdio + Streamable HTTP transport (#381)
docgen	Build tooling — auto-generates api-reference.md from route registry (publish = false)

Components

Component	Technology	Detail
Language	Rust (no unsafe)	Memory-safe, fast, single binary
Vector Index	usearch + RwLock	Persistent HNSW, concurrent reads via RwLock
Full-Text Search	SQLite FTS5	Built-in, auto-created, phrase + token-OR fallback
Hybrid Search	RRF (k=60)	Merges vector + FTS5 via Reciprocal Rank Fusion
Fusion (default)	weighted RRF	Since 0.16.0: fuses the vector and hybrid rankings (#1123)
Storage	SQLite (rusqlite)	Embedded, zero-config, battle-tested
Embedding	EmbeddingGemma Q4 ONNX	768d vectors, multilingual, downloaded on first run
Namespaces	SQLite column	Multi-agent isolation, zero overhead
Tiered Memory	Access tracking	Hot/Warm/Cold scoring boost
CLI	clap	Standard Rust CLI

How It Works

1. remember — Text is embedded into a 768d vector via ONNX → stored in SQLite + indexed in usearch
2. recall — Query is embedded → usearch finds nearest neighbors + FTS5 finds keyword matches → RRF merges both result sets → hot memories get +0.1 score boost → returns ranked results
3. search — SQLite FTS5 keyword search (phrase match + token-OR fallback, scoped to namespace)
4. forget — Incremental delete from usearch + SQLite (no rebuild)
5. Everything lives in ~/.codecora/uteke/ — fully local, fully yours

Data Flow

Write Path (remember)

Content → ONNX embed (Mutex) → 768d vector
                                ↓
                    SQLite INSERT (metadata + vector)
                                ↓
                    usearch INSERT (RwLock write)
                                ↓
                    usearch SAVE (atomic: .tmp + rename)

SQLite-first dual-write: metadata persisted before index. If index write fails, data is still in SQLite and can be recovered via uteke repair.

Read Path (recall)

Query → Recall Cache check (TTL 5min, LRU 256)
         ↓ miss
       ONNX embed (Mutex) → 768d vector
                                ↓
              ┌─────────────────┴──────────────────┐
              ↓                                     ↓
     usearch SEARCH (RwLock read)         FTS5 SEARCH (SQLite)
     → top-k vectors + scores             → top-k rows + BM25
              ↓                                     ↓
              └───────────── RRF Merge ─────────────┘
                            ↓
                    Tier boost (+0.1 hot)
                            ↓
                    Entity/category filter (#667)
                    (pushed into core recall loop)
                            ↓
                    Ranked results

Delete Path (forget)

ID → usearch REMOVE (RwLock write, acquired first)
                           ↓
              SQLite DELETE (by specific ID)
                           ↓
              usearch SAVE (atomic)

Index lock acquired before SQLite delete — narrows the inconsistency window.

Extract / Import / Export (v0.7.0)

Endpoint	Method	Auth	Description
/extract	POST	Write	LLM fact extraction + auto-store (1MB body limit)
/import	POST	Write	JSONL import with re-embedding (5MB body limit)
/export	GET	Read	JSONL export (optional ?namespace= filter)

Maintenance & Monitoring (v0.7.0)

Endpoint	Method	Auth	Description
/prune	POST	Write	TTL-based deprecated memory cleanup
/consolidate	POST	Write	Near-duplicate merging
/aging	POST	Write	Memory lifecycle: status, preview, cleanup
/importance	POST	Write	Recalculate importance scores
/orphans	POST	Read	Find disconnected, low-importance memories
/rebuild-backlinks	POST	Write	Rebuild referenced_by from forward edges

Document Update (v0.7.0)

Endpoint	Method	Auth	Description
/doc/update	POST	Write	Partial document update with chunk rebuild

Memory Mutation (v0.8.0)

Endpoint	Method	Auth	Description
/memory	PUT	Write	Partial memory update (content, tags, metadata, importance, pinned, memory_type)
/memory/pin	POST	Write	Pin or unpin a memory by ID
/memory/importance	POST	Write	Set importance score (0.0–1.0) for a memory

Room ↔ Document Junction (v0.8.0)

Endpoint	Method	Auth	Description
/room/document/list	POST	Read	List document slugs linked to a room
/room/document/add	PUT	Write	Link a document to a room
/room/document/remove	DELETE	Write	Unlink a document from a room
/doc/room/list	POST	Read	List rooms linked to a document

Cross-Entity References (v0.8.0)

Endpoint	Method	Auth	Description
/memory/doc-refs	POST	Read	Get document slugs referenced by a memory
/doc/mem-refs	POST	Read	Get memory IDs that reference a document

Key Design Decisions

RwLock for Vector Index (not Mutex)

Read-heavy workload: recall/search operations far outnumber remember/forget. Multiple concurrent recalls share a read lock. Embedder remains Mutex because ONNX tokenizer requires &mut self internally.

Cross-Process File Lock (#543)

The RwLock provides intra-process concurrency only (multiple threads within a single uteke-serve process). For cross-process safety (multiple uteke-serve instances or CLI + server sharing the same store), a file-based lock (fs2 crate) guards the SQLite database and usearch index files. The lock is acquired on store open and held until the process exits or the store is dropped. This prevents concurrent writes from corrupting the index or database when multiple processes access the same ~/.codecora/uteke/ directory.

FTS5 + Vector Hybrid via RRF

Two systems with incompatible score scales (cosine 0..1 vs BM25 unbounded). Reciprocal Rank Fusion solves this by ranking based on position, not score magnitude. k=60 is the standard literature value.

Graph-Augmented Reranking (#378)

The graph recall strategy layers graph signals on top of the RRF result. compute_graph_signals() issues a single batched query over memory_edges (derived from [[slug]] / @tag / ^id auto-wiring, #346) and computes per-memory density/authority counts. rerank_with_graph() then applies an additive, log-scaled boost (ln(1+x) * weight) so well-connected memories drift upward while isolated ones are untouched. The boost saturates quickly (going 1→10 edges ≈ 100→1000 in lift), preventing hub dominance. The recall cache is strategy-keyed, so graph entries never collide with hybrid/ vector/fts5.

Atomic File Writes

All critical file I/O uses the .tmp + rename pattern. On POSIX filesystems, rename is atomic — a crash mid-write never leaves a corrupt file, only the old version.

Schema Versioning

Integer counter in schema_version table. Migrations run automatically on upgrade. Currently at v15. Schema history: v4 rooms, v5 memory_tags junction, v6 content_type column, v7 knowledge graph, v8 memory_edges + slug (#346), v9 timeline_events table (#347), v10 source + source_type (#348), v11 documents + document_chunks (#406), v12 hierarchy (parent_id on documents, room tables added to SCHEMA constant), v13 global documents (namespace deprecated, author column, slug uniqueness), v14 memory_type added to FTS5 index (#662), v15 room_documents junction table for room↔document linking (#689). Zero data loss guaranteed.

Rooms

Rooms group related memories with author attribution. Stored in rooms and room_memories tables (schema v4). Semantic room recall over-fetches via recall_hybrid(), then post-filters to room memory IDs. Room summaries use tag co-occurrence clustering — no LLM call required. Room documents group memories by memory_type into sections (Decisions, Facts, Procedures, etc).

Rooms can also be linked to documents via the room_documents junction table (schema v15, #689), enabling bidirectional room↔document associations. Documents reference memories via [[doc-slug]] wikilinks, which are auto-resolved to cross-entity edges (#691).

Memory Types (#349)

Every memory has a memory_type: fact, procedure, preference, decision, context, note, insight, reference, event. Type is set explicitly via --type or auto-inferred from content patterns (questions → context, URLs → reference, dates → event, etc.).

Timeline Events (#347)

The timeline_events table (schema v9) records an append-only audit log per memory: creation, updates, type changes, supersession, edge additions, pin/unpin. Events store metadata as JSON. Consolidation events attach to the OLD memory with replaced_by metadata.

Salience + Recency (#352)

Dual-axis recall boost applied after the RRF merge:

• Salience: higher-scored by type weight (decision > insight > fact > note)
• Recency: exponential decay exp(-age/τ) where τ is a per-type time constant

Both are enabled by default (weight 0.1). Use --no-salience / --no-recency to disable. Previously opt-in (v0.7.3), now opt-out (#721). Config leaks are prevented by setting/resetting config around each query.

Dream Cycle (#353)

The dream command runs a coordinated maintenance pipeline: lint → backlinks → dedup → contradict → orphans → compact → verify. Phase 4 (Contradict) scans for contradictory memories via tag overlap + embedding divergence (#720). Phases run in canonical dependency order regardless of user-supplied order. Individual phase errors are recorded but don't abort the pipeline. Namespace-scoped phases (lint, dedup, contradict, orphans, compact) only affect the target namespace; global phases (backlinks, verify) run across all namespaces.

Citation & Source Attribution (#348)

The source and source_type columns (schema v10) track where a memory came from. Source types: user, url, file, import, derived, system, unknown. Set via --source / --source-type on remember or set_source() post-insert.

MCP Transport (#381)

The uteke-mcp crate provides a shared JSON-RPC handler used by both:

• stdio binary (uteke-mcp) — for local agents (Claude Desktop, Cursor)
• HTTP endpoint (POST /mcp on uteke-serve) — for remote MCP clients

Protocol version: 2025-06-18 (Streamable HTTP spec). 1 MiB body limit enforced on HTTP endpoint. JSON-RPC 2.0 strict compliance (v0.6.7, #573/#576): tagged union responses, no notification response, Claude Code compatible.

Performance

Mode	Recall	Setup
Library (Rust)	~45ms	In-process, no startup
Server (HTTP)	~42ms	One-time ~2s init
CLI (binary)	~3s	Per-invocation (model load)

CLI vs Server

Metric	CLI (cold)	Server (warm)	Speedup
Insert 100	~316s (0.3/s)	7.7s (13/s)	41x
Recall (avg)	3,158ms	42ms	75x
Search (avg)	3,158ms	9ms	367x

Scaling (warm server)

Data Size	Recall (avg)	Search (avg)
100 memories	42ms	9ms
1,000 memories	49ms	13ms
10,000 memories*	~55ms (est.)	~20ms (est.)

*10K estimated — HNSW vector search is O(log n)

Benchmarked on Oracle Cloud ARM (Ampere Altra), CPU-only, no GPU.

File Structure

~/.codecora/uteke/
├── uteke.db                    # SQLite (memories + metadata + FTS5)
├── uteke_index.usearch         # Persistent HNSW vector index
├── uteke_index.keys            # Index key mapping (atomic save)
├── embeddinggemma-q4/           # Local ONNX embedding model (~188MB)
│   └── onnx/                    # model_q4.onnx + model_q4.onnx_data
└── logs/
    ├── uteke.log               # Current log
    └── uteke.log.YYYY-MM-DD    # Rotated logs

In Docker, the data directory defaults to /data (set via UTEKE_HOME). The slim image variant does not bundle the embedding model — mount the model directory separately:

docker run -d --name uteke \
  -v uteke-data:/data \
  -v ./models/embeddinggemma-q4:/data/embeddinggemma-q4 \
  ghcr.io/codecoradev/uteke:slim

Known Limitations

Limitation	Status	Detail
usearch ef not configurable	External	usearch v2.25.3 Rust bindings don't expose ef in search()
Embedder requires Mutex	Architectural	ONNX tokenizer internally uses &mut self
Consolidate is O(n²)	Algorithm	Pairwise cosine — slow above 1K memories
FTS5-only score is placeholder	Design	BM25 can't normalize to 0..1; actual ranking via RRF

Document Engine (#406)

Full markdown content → SQLite (documents table), chunked summaries → embeddings (document_chunks).

Schema (v11)

-- Full documents with unlimited content
documents: id, slug, title, content, namespace, tags, metadata,
           version, content_type, created_at, updated_at

-- Chunked sections with per-chunk embeddings
document_chunks: id, document_id, chunk_index, heading, content,
                 embedding BLOB, char_start, char_end, tags

Chunking Pipeline

1. Document content → chunk_markdown() (#405) splits by headings
2. Each chunk → embedder.embed() creates section-level embedding
3. Chunks stored with heading, content, char offsets, embedding BLOB

Chunk size derived from embedder.max_seq_len() (#407): ~4 chars per token.

• ONNX (256 tokens): 1,024 chars/chunk
• OpenAI (8191 tokens): 32,764 chars/chunk

Search Paths

1. Direct ID/slug → SQLite PK → full content
2. Semantic → embed query → ANN chunk search → resolve to document
3. FTS5 → full text across all document content
4. Graph → BFS from entity → connected documents

Cosine Auto-Linking (#401)

After every remember(), the vector index is searched for the top-20 most similar memories:

• Cosine ≥ 0.80 → similar_to edge
• Cosine ≥ 0.92 → possible_duplicate edge

Edges are namespace-scoped (no cross-namespace links). Best-effort: errors logged, never fail remember().

Graph API (#408)

GET /graph returns all graph nodes, edges, and stats as JSON for visualization clients:

{
  "nodes": [{ "id": "...", "label": "...", "entity_type": "..." }],
  "edges": [{ "source_id": "...", "target_id": "...", "relation": "..." }],
  "stats": { "node_count": N, "edge_count": N }
}