Vestige is a local-first deterministic operating system, execution safety kernel, and zero-vector causal memory engine for autonomous AI agents built in Rust.
Unlike probabilistic LLMs and vector RAG databases that rely on floating-point similarity without tracking causality, Vestige wraps AI agents in a local Rust binary. It enforces cryptographic execution gates, bi-temporal state tracking, and zero-vector causal retrieval with $0 API cost and no GPU overhead.
🚀 Live WebGPU Companion (Zero Install): https://stara-companion.pages.dev/?v=pro-reset
npx @vestige/init --guardDetects installed AI agents, configures the MCP server, and installs the Operator safety guard.
{
"mcpServers": {
"vestige": { "command": "npx", "args": ["-y", "vestige-mcp-server"] }
}
}(Or in Antigravity: agy plugin add vestige)
Download pre-built standalone binaries directly from GitHub Releases, or install via shell:
curl -fsSL https://raw.githubusercontent.com/samvallad33/vestige/main/scripts/install.sh | sh| Capability | Vestige (samvallad33/vestige) |
Vector RAG Memory (Mem0, Zep, Letta) | Policy Wrappers (Microsoft Agent OS, Kern) |
|---|---|---|---|
| System Category | Deterministic Agent OS & Causal Kernel | Probabilistic Vector Memory Store | Stateless Middleware / Policy Firewall |
| Retrieval Engine | Zero-Vector Causal Math (RSB v4, Rank-1 seating, FSRS-6) | Dense Embeddings + ANN Cosine Similarity | None / External RAG Dependency |
| Compute & API Cost | $0 / CPU-Only (verified on 2011 Core i5-2520M) | GPU / Cloud Embedding & Reranker APIs | Low CPU (Policy Rules Only) |
| State & Audit Ledger | Strata: Append-only, signed, BLAKE3 hash-chained, bi-temporal | Mutable DB rows / overwrites | Append-only audit logs (no causal memory graph) |
| Execution Governance | Out-of-Band Operator Gates (--guard) + 15-min owner permits |
None (relies on agent self-restraint) | Pre-execution policy & sandbox checks |
| Root-Cause Tracing | causal_walk & vestige prove (automated git bisect) |
Not supported | Log inspection only |
| Multi-Agent Sync | Reflex Context Manifest + Stale-Belief Push | Shared vector namespace (stale drift risk) | Per-process isolation |
- Strata (Cryptographic Causal Ledger): Append-only, signed, BLAKE3 hash-chained log providing bi-temporal validity bounds (
valid_from_ms,valid_until_ms). Writes return receipts; new claims supersede older ones while historical facts remain immutable. - Operator Gates & Reflex Hooks (
--guard): Out-of-band execution governance requiring 15-minute single-use owner permits for high-risk actions, logging every block and allowance to a dedicated ledger. - Zero-Vector Causal Retrieval (RSB v4 + Rank-1 Seating + FSRS-6): Structural provenance engine using Retroactive Salience Backfill, deterministic slot-1 seating rules, and FSRS-6 memory decay instead of embeddings. Cause never looks like the symptom — a later turn raises earlier relevant context even with zero shared words.
- Causal Root-Cause Proofs (
causal_walk&vestige prove): Traces upstream failure and commit edges, running deterministic git-bisect verification against your own tests with receipted runs. - Multi-Agent Context Manifest & Compaction Survival: Manages context token budgets, pushes live stale-belief revocations across subagents, and restores decisions across context compactions.
Benchmarked using the local retrieval_bench binary on an Intel Core i5-2520M with 8 GB RAM, CPU only (no embeddings, no LLM, no GPU, $0 API costs):
| Benchmark | Split / Slice | Level | Metric | Score |
|---|---|---|---|---|
| LoCoMo-10 | Blind Convs 6–10 (n = 985, out-of-sample) |
Session / Turn | Sess R@5 / Turn Exact R@5 / Turn R@10 | 93.71% / 80.41% / 88.02% (~59ms/q) |
| LoCoMo-10 | All 10 Convs (n = 1,982) |
Session / Turn | Sess R@5 / Turn Exact R@5 / Turn Soft R@5 | 94.25% / 81.84% / 84.71% |
| LongMemEval-S | Official (470 non-abstention) |
Session | Any@1 / Any@5 / Any@10 / CappedAll@5 | 85.74% / 96.17% / 99.57% / 92.08% |
| BEAM 100K | 355 scored questions (2,127 sub-sessions) |
Session / Turn | Pkt Sess Any@5 / Pkt Turn Any@5 / Temporal Any@5 | 78.59% / 70.42% / 100.0% |
| MemoryAgentBench | QA2 421K (100 questions) |
Document | All@5 | 85/100 (85.00%) |
Reproduce locally:
bash benchmarks/run_all.shFull tables and methodology are in benchmarks/README.md.
Built by Sam Valladares (@samvallad33). Licensed under AGPL-3.0.