OSpipe: Quantum-Enhanced Personal AI Memory
OSpipe integrates Screenpipe — an open-source, local-first desktop recording and AI memory system — with the RuVector ecosystem of 70+ Rust crates and 50+ npm packages. The result is a privacy-preserving personal AI memory system with semantic vector search, knowledge graphs, temporal analysis, and cross-platform deployment.
Core Idea
Screenpipe captures everything you see, hear, and do on your desktop. OSpipe replaces its SQLite + FTS5 storage with RuVector's SIMD-accelerated HNSW vector database, adding:
- Semantic Search — Find content by meaning, not just keywords (61us p50 latency)
- Knowledge Graph — Relationships between apps, people, topics, meetings
- Temporal Analysis — Delta-behavior tracking for change detection
- PII Safety Gate — Credit card, SSN, email redaction before storage
- Frame Deduplication — Cosine similarity sliding window eliminates duplicates
- Cross-Platform — Native (Windows, macOS, Linux) + WASM (browser)
Features
| Feature |
Description |
Status |
| Rust Core Library |
7 modules: capture, storage, search, pipeline, safety, config, wasm |
Done |
| WASM Bindings |
wasm-bindgen exports for browser deployment |
Done |
| TypeScript SDK |
@ruvector/ospipe with SSE streaming, hybrid search |
Done |
| PII Safety Gate |
Credit card, SSN, email detection + redaction |
Done |
| Frame Deduplication |
Cosine similarity sliding window |
Done |
| Query Router |
Heuristic routing to Semantic/Keyword/Graph/Temporal/Hybrid |
Done |
| Hybrid Search |
Semantic + keyword score fusion |
Done |
| Configurable Quantization |
4-tier age-based (f32 → int8 → product → binary) |
Done |
| 56 Tests |
24 unit + 32 integration, all passing |
Done |
| Cross-Platform Builds |
Linux x64, Windows GNU, WASM32 verified |
Done |
| Knowledge Graph Integration |
ruvector-graph for Cypher queries |
Planned (Phase 2) |
| Attention Streaming |
Multi-head attention for content prioritization |
Planned (Phase 3) |
| GNN Search Improvement |
Learned retrieval improvement over time |
Planned (Phase 3) |
| Quantum Search |
Grover-inspired amplitude amplification |
Planned (Phase 4) |
| On-Device LLM |
ruvllm GGUF inference, fully offline |
Planned (Phase 5) |
Comparison: Screenpipe vs OSpipe
| Aspect |
Screenpipe (FTS5) |
OSpipe (RuVector) |
| Search Type |
Keyword only (FTS5) |
Semantic + Keyword + Graph + Temporal + Hybrid |
| Search Latency |
~1ms |
61us (HNSW p50) |
| Content Relations |
None |
Knowledge Graph with Cypher |
| Temporal Analysis |
Basic SQL queries |
Delta-behavior causal tracking |
| PII Protection |
Basic |
Credit card, SSN, email redaction |
| Deduplication |
None |
Cosine similarity gate (configurable threshold) |
| Browser Support |
None |
WASM (11.8KB micro to 350KB full) |
| Quantization |
None |
4-tier age-based (100% → 3% memory) |
| Privacy |
Local-first |
Local-first + automatic PII redaction |
| Embedding Model |
External (Ollama/OpenAI) |
On-device ONNX + GGUF (planned) |
Architecture
Screenpipe Capture → Safety Gate → Dedup → Embed → HNSW VectorStore
|
Query Router ←-------------------+
├── Semantic (HNSW cosine)
├── Keyword (FTS5 compat)
├── Graph (Cypher via ruvector-graph)
├── Temporal (delta-core replay)
└── Hybrid (score fusion)
Quick Start
Rust:
use ospipe::pipeline::ingestion::IngestionPipeline;
use ospipe::config::OsPipeConfig;
let pipeline = IngestionPipeline::new(OsPipeConfig::default())?;
// Ingest captured frames, search by semantic meaning
TypeScript:
import { OsPipe } from "@ruvector/ospipe";
const client = new OsPipe();
const results = await client.queryRuVector("what did I discuss about auth?");WASM:
const pipe = new OsPipeWasm(384);
pipe.insert("f1", pipe.embed_text("meeting notes"), '{}', Date.now());
const results = pipe.search(pipe.embed_text("notes"), 5);RuVector Crate Integration
| Crate |
Purpose |
Phase |
ruvector-core |
HNSW vector storage + SIMD |
1 (Done) |
ruvector-filter |
Metadata filtering |
1 (Done) |
ruvector-cluster |
Frame deduplication |
1 (Done) |
ruvector-delta-core |
Change tracking |
1 (Done) |
ruvector-router-core |
Query classification |
1 (Done) |
cognitum-gate-kernel |
AI safety decisions |
1 (Done) |
ruvector-graph |
Knowledge graph |
2 |
ruvector-attention |
Content prioritization |
3 |
ruvector-gnn |
Learned retrieval |
3 |
ruqu-algorithms |
Quantum-enhanced search |
4 |
ruvllm |
On-device LLM |
5 |
Implementation Details
- Branch:
feat/ospipe-screenpipe-integration
- Crate:
examples/OSpipe/ (14 Rust source files)
- TypeScript SDK:
npm/packages/ospipe/
- WASM Package:
npm/packages/ospipe-wasm/
- ADR:
examples/OSpipe/ADR-OSpipe-screenpipe-integration.md (1,986 lines)
- Tests: 56 passing (24 unit + 32 integration)
- Build Targets: Linux x64, Windows x64 GNU, wasm32-unknown-unknown
Configuration Reference
CaptureConfig
| Parameter |
Default |
Description |
| fps |
1.0 |
Screen capture frames per second |
| audio_chunk_secs |
30 |
Duration of audio recording chunks |
| excluded_apps |
["1Password", "Keychain Access"] |
Apps to skip |
| skip_private_windows |
true |
Skip incognito/private windows |
StorageConfig
| Parameter |
Default |
Description |
| embedding_dim |
384 |
Vector dimensions |
| hnsw_m |
32 |
Max connections per HNSW layer |
| hnsw_ef_construction |
200 |
Build-time search width |
| hnsw_ef_search |
100 |
Query-time search width |
| dedup_threshold |
0.95 |
Cosine similarity dedup cutoff |
SearchConfig
| Parameter |
Default |
Description |
| default_k |
10 |
Results per query |
| hybrid_weight |
0.7 |
Semantic vs keyword blend |
| mmr_lambda |
0.5 |
Diversity vs relevance tradeoff |
| rerank_enabled |
false |
Enable result reranking |
SafetyConfig
| Parameter |
Default |
Description |
| pii_detection |
true |
Enable PII detection |
| credit_card_redaction |
true |
Redact credit card numbers |
| ssn_redaction |
true |
Redact SSNs |
| custom_patterns |
[] |
Custom deny patterns |
Quantization Tiers
| Age |
Method |
Memory vs f32 |
| 0h |
None (f32) |
100% |
| 24h |
Scalar (int8) |
25% |
| 1 week |
Product |
~6% |
| 30 days |
Binary |
3% |
Safety Gate Details
The safety gate inspects content before storage:
| PII Type |
Pattern |
Replacement |
| Credit Card |
13-16 digit sequences |
[CC_REDACTED] |
| SSN |
XXX-XX-XXXX |
[SSN_REDACTED] |
| Email |
user@domain.tld |
[EMAIL_REDACTED] |
| Custom |
Configurable patterns |
[DENY] |
Decisions: Allow → store as-is | AllowRedacted → store cleaned | Deny → reject entirely
WASM Deployment Tiers
| Tier |
Size |
Capabilities |
| Micro |
11.8KB |
Embedding + search only |
| Standard |
225KB |
Full pipeline |
| Full |
350KB |
+ Dedup + safety |
| AI |
2.5MB |
+ On-device inference |
