AI memory that stays
on your machine.
VeritasMemoria is a structured, auditable memory layer for AI-assisted workflows, built on a geometry-grounded knowledge graph that separates durable governance from active work. Every record is cryptographically signed and hash-chained. No cloud. No accounts. No telemetry.
Governance at the core · contested claims at the edge
Built different by design
Most memory systems trade transparency for convenience. VM trades neither — knowledge is organized by epistemic role and maturity, governed by explicit rules, and every change leaves a verifiable trace.
Local-First
All data stays on the machine running it. Nothing is sent to any server — not even to check for updates. Your memories are yours.
Tamper-Evident
Records are cryptographically signed and chained into a Merkle-verifiable audit trail. Retired memories move to an append-only archive rather than disappearing — you can always trace what changed, when, and why.
Human-in-the-Loop
The system cannot commit or retire memories without passing configurable oversight gates. You stay in control of what gets kept.
LLM-Agnostic
Works with OpenAI, Anthropic, Ollama, or any OpenAI-compatible endpoint. Runs fully offline in copy-paste mode with no LLM required.
Hybrid Retrieval
Combines dense vector embeddings, keyword indexing, and graph traversal. Finds the right memory whether you remember the exact phrasing, a related concept, or just that it connects to something else you're looking at.
Transparent Storage
SQLite by default, a single file, no proprietary format and no vendor lock-in. It isn't meant to be hand-edited though: the hash chain is deliberately fragile to any change made outside VM, so tampering is detectable rather than invisible. Read it, back it up, move the file. Don't edit it directly.
Isolated Matters, Shared Governance
Each project, case, or workstream lives in its own isolated space. Nothing crosses between them directly — shared policy and oversight live in one governed layer that every workstream defers to.
Regulatory-Grounded Permissions
Every action an agent can take is checked against layered permission rules tied to domain-specific compliance concerns, legal, medical, or financial, so higher-level policy can only restrict, never expand, what's allowed. Built to meet EU AI Act high-risk system obligations natively: audit trails, human oversight, and traceability aren't bolted on afterward, they're how the system works by default.
Geometry-grounded, not just embedded
The belief graph isn't a bag of vectors ranked by cosine similarity. Distance, curvature, and convergence are real signals the system acts on, not decoration on top of a standard database.
Hyperbolic Belief Geometry
Every memory occupies a real position in a curved graph, placed by how settled and how central it is. Governance sits at the core, load-bearing and hard to move. Contested, provisional material sits out toward the edge, where it belongs.
Knows When Reasoning Has Converged
Most systems answer as soon as they retrieve something plausible. VM tracks a spectral stability signal across the graph and won't let a decision commit until reasoning has actually settled, not just until something matched the query.
Hunts Its Own Blind Spots
A dedicated adversarial search runs against the belief graph looking for states that look internally consistent but actually reason in circles. Anything it finds gets flagged for human review before it can be acted on.
Full Provenance, Any Point in Time
Every write is a dated deposit. Every correction is recorded as an event that disturbs the record rather than erasing it. Ask what the system believed as of any past date and it reconstructs that exact state.
Nothing Is Truly Deleted
Removed memories move to an append-only shadow archive and stay there indefinitely, not until a retention window closes. For regulated work, VM also retains the model's actual reasoning trace, every tool call and intermediate step, not just the final answer, so you can audit how a conclusion was reached.
Simple architecture, serious guarantees
Under the hood, knowledge is organized by role and maturity rather than dumped into one flat index — governance stays stable while active work and evidence can shift, and contested information is treated differently from settled information.
You write a memory
Paste text, upload a document, or let an agent push context. VM assigns it to working memory.
VM gates and signs it
Coherence and oversight checks run. On approval the record is HMAC-signed and appended to the hash chain.
Retrieve it later
Hybrid semantic, keyword, and graph search surface the right context instantly. Pass it to any LLM or read it yourself.
Request beta access
VeritasMemoria is currently in private beta and isn't publicly available yet. Tell us a bit about your use case and we'll follow up by email.
Requests are sent directly to jessica@templetsolutions.com. We typically respond within a few business days.
Works with what you already use
Switch providers in a single line of config. Or skip the LLM entirely and run in copy-paste mode.
Your data. Your machine. Your call.
VeritasMemoria calibrates to your domain's ontology out of the box, then learns only what you explicitly give it from there. No defaults phoning home. No model training on your memories.