Prime-Silo is a browser-first agent runtime that replaces ephemeral chat context with a permanent, auditable tri-graph cognitive mesh. Run state-of-the-art local models with cryptographic governance and zero cloud leakage.
Prime-Silo is not a bundle of features — it is a small set of load-bearing primitives that every workflow composes. Build them once, honestly, and each new capability costs less than the last.
Memo-Ray X-rays every agent session — across Claude Code, Antigravity and opencode — into one queryable timeline store instead of write-only logs.
Unlocks: Long-horizon anything: you cannot collate what you never kept.
One router in front of LM Studio, Lemonade and Ollama with an offline guard; models are env-profile driven, never hardcoded.
Unlocks: Bulk cognition at electricity prices — and eval sweeps by swapping one variable.
A single Source → Concept graph schema shared by documents, code and memory — derived deterministically wherever the entities are already known.
Unlocks: Cited knowledge with no hallucination surface where it matters.
HMAC-SHA256 on every manifest; an append-only ledger that records failures as plainly as successes; CLP lineage end to end.
Unlocks: Trust: every claim traceable, every run auditable, every gap visible.
Phase checkpoints, per-item caches, reconcile-on-restart and halt-on-fault — a killed run resumes exactly where it stopped.
Unlocks: Multi-day workflows on consumer hardware that survive sleeps, crashes and wedged engines.
Isolated, disposable workspaces under one portable home — the whole runtime ships zero-install on six platforms.
Unlocks: Clean experiments, honest A/B corpora, and the same stack on any machine.
Switch between live command simulations to watch how Prime-Silo processes knowledge, synthesizes graphs, and verifies local memory integrity.
Traditional AI agents suffer from amnesia. They discard context after every session or dump unstructured chat transcripts into flat databases that degrade over time.
Sending sensitive IP, proprietary codebases, and executive thinking through metered cloud APIs exposes your organization to ongoing privacy hazards and recurring operational spend.
Unverified vector search retrieves fragmented paragraphs out of context, forcing models to stitch together hallucinated answers without clear source provenance.
When autonomous agents execute tools or modify files without cryptographic signatures and verifiable execution lineage, security audits become impossible.
Language models are probabilistic reasoning engines—they excel at synthesis but cannot guarantee determinism. Prime-Silo draws a strict boundary between what the model generates and what the underlying runtime guarantees.
Never trust an agent whose memory you cannot inspect, verify, and own end-to-end.
Unified multi-layer graph connecting documents, Tree-Sitter code ASTs, and Memo-Ray session memories.
High-throughput local knowledge distillation converting thousands of past interactions into structured hubs.
Bronze-silver-gold data refinement pipelines with end-to-end Data Lineage Protocol (CLP) tracking.
Integrated desktop command center for live agent observation, process inspection, and resource management.
Native integration with LM Studio, Lemonade, and Ollama with hardware fallback and strict offline guards.
Intelligent task scheduling balancing local GPU/NPU execution with optional secure proxy endpoints.
Full support for Model Context Protocol servers and Agent-to-Agent communication channels.
HMAC-signed execution records ensuring audit readiness and verifiable provenance for every operation.
Six specialized runtime modes for developer workflows, architectural review, and autonomous supervision.
Instant state snapshots allowing time travel, branch exploration, and safe rollback of agent edits.
Cross-platform native desktop runtime for Windows, macOS, and Linux with self-updating supervisor.
Ingest architectural diagrams, UI mockups, and figures directly into the knowledge graph.
Built-in accessibility suite featuring OpenDyslexic typography, bionic reading, and calm motion controls.
Ledger-sourced OpenLineage DAG with per-artifact ontology, step-through replay, and a workspace-grouped execution register.
Label and relocate sensitive sessions into an isolated workspace — files, graph nodes, and vectors together — journalled, reversible, and leak-gated.
Spec-compliant OpenLineage RunEvents written to an integrity-hashed governance ledger and rendered live - progress, earned ETA, execution register groupable by output, commit, or workspace. No lineage server required.
One CLI builds an enterprise SAD from the dual graph: diagrams-as-code from graph truth, recursive index planning, per-section citation gates, robustness catalogs, and a PDF with every diagram realized.
Unlike ephemeral chatbots that reset to zero every morning, Prime-Silo compounds institutional knowledge over time. Every task executed enriches the tri-graph mesh; every LONGVIEW run synthesizes new patterns; and every signed manifest strengthens your audit trail. The result is a sovereign engineering partner that is greater than the sum of its parts.
Interactive audio overviews, deep-dive podcasts, and executive briefings generated locally from your LONGVIEW knowledge graph.
Unified execution stream with live real-time NPU, GPU, and VRAM telemetry visualization.
Distributed multi-device worker clustering across LAN hardware for high-throughput repository indexing.
Curated, verified community skill sets and deterministic extension packages.
Adaptive compute scheduling that automatically scales between low-power CPU, local eGPU, and self-hosted clusters.
One-click generation of tamper-evident compliance archives for SOC2 and internal IP audits.
Automated detection of conflicting assertions and outdated technical concepts across long-lived agent memories.
Autonomous background maintenance tasks that clean up stale artifacts and draft documentation updates.
No. Prime-Silo runs entirely on your local machine by default. All knowledge graphs, session transcripts, and files remain on your disk under your direct control. Outbound network requests occur only if you explicitly configure an external cloud provider.
Prime-Silo runs comfortably on Apple Silicon (M1/M2/M3/M4 with 16GB+ unified memory) or Windows/Linux systems with an NVIDIA or AMD GPU (8GB+ VRAM recommended for 7B-12B models). For lower-powered hardware, it can orchestrate lightweight quantizations or LAN servers.
Prime-Silo connects seamlessly to LM Studio, Lemonade, Ollama, and standard OpenAI-compatible local endpoints. You can swap models per task without altering your underlying knowledge graph.
Standard RAG simply splits text into vector chunks. Prime-Silo builds a deterministic semantic graph linking documentation, Tree-Sitter syntax trees (ASTs), and historical agent memories via explicit CORRELATES_WITH relationships—enabling precise structural reasoning.
Yes. Prime-Silo core modules and browser runtime are open-source under the MIT License, ensuring transparency and freedom from vendor lock-in.
Yes. Every tool execution and file modification generates an HMAC-signed manifest and Data Lineage Protocol (CLP) record, creating an unforgeable chain of custody for compliance reviews.
Download the desktop app or explore our source code on GitHub. Join us in building AI systems where you own the memory.
Direct Contact: binary16.primesilo@gmail.com