Recall, remember, edit, and explore your memory the way your AI client does — using the exact same operations, running entirely on your device. No internet. No cloud hop. No third-party model seeing your data. Just you, your cortex, and a deterministic graph that never lies.
The problem with cloud memory
Problem 1
Cloud AI memory services store your notes, corrections, and personal context on infrastructure you don't control, under retention policies you didn't write.
Graphnosis
Your cortex is an encrypted file on your disk. No account, no sync, no one else's cloud. The passphrase never leaves your device.
Problem 2
Cloud memory managers give your AI a summary. You never see which nodes were recalled, how they scored, or why a fact was included or dropped.
Graphnosis
MemoryStudio shows you the full subgraph output — the exact structured context your AI client receives, with node types, confidence scores, and source provenance.
Problem 3
Cloud memory updates are opaque rewrites. There's no diff, no audit trail, no way to see what changed or roll back a bad correction.
Graphnosis
Every correction goes through a diff-review step. You see before and after. You approve or reject. The op-log records the full lineage — forever.
What MemoryStudio does
The same recall, remember, edit, and explore loop your AI client runs via MCP — available directly in the app, offline, without sending a single byte to a server. Seed your engrams with a Memory Kit →
Type a query in any language. Read the exact structured subgraph your AI client would receive via MCP — nodes, edges, confidence scores, and source provenance. No surprises, no black box.
Deterministic · Same query = same result
When recall returns thin results, Dig Deeper runs three expansion stages: source-filename search, cross-engram entity hop, and GNN graph expansion. Finds the memory that bare recall misses.
Multi-stage · Provenance breakdown included
Write a note. The app suggests the right engram based on content similarity, then checks for near-duplicates before saving. Your cortex stays clean, structured, and non-redundant.
Auto-routed · Duplicate check built-in
Describe a correction in plain language. The app surfaces the affected memory, proposes a before/after diff, and waits for your approval. With Ollama running locally, the LLM parses your intent — without sending anything to a cloud model.
Diff-reviewed · Op-log auditable
A local graph neural network scans embedding similarity, shared entities, and graph topology to predict connections your explicit notes never made. Explore the hidden structure of your memory — the links that exist but haven’t been written yet.
Local inference · Opt-in overlay
Before saving any note, a cosine similarity scan checks your entire cortex for near-identical content. A warning surfaces the best match so you can decide: save as new, edit instead, or discard.
No LLM required · Threshold-configurable
How it works
Ask in any language. Graphnosis searches your local encrypted graph using BGE multilingual embeddings — no API call, no cloud hop. Median: 75 ms.
Read the exact structured context your AI client receives via MCP — nodes, edge weights, confidence scores, source provenance. The same view, in your hands, without an AI intermediary.
Save a note, fix a fact, explore a GNN-predicted connection —
every action writes to your local .gai file and nothing else.
75 ms
Median recall latency
76.4%
LongMemEval accuracy
$0
Per-query cloud cost
100%
Data stays on your device
The same query against the same cortex always returns the same result — no randomness, no model drift, no API non-determinism. Your local BGE model runs the search. When you enable Ollama, the LLM runs on your GPU too. The only thing that travels is what you choose to share — and Graphnosis was built to make that choice explicit, not accidental.
Encrypted at rest
XChaCha20-Poly1305
75 ms recall
Local BGE embeddings
Local LLM
Ollama, optional
No telemetry
No analytics, ever
Always running
MemoryStudio surfaces two background engines that keep your graph healthy and connected. They don’t need a prompt. They don’t need internet. They run while you do other things.
No LLM required · No internet · Byte-reproducible
“This is what your memory would do naturally if it had a maintenance crew. Pure graph math — no AI, no API, no guessing.”
Memory health gauge
Vitality score 0–100 based on node connectivity, recency, and redundancy. Know at a glance whether your cortex is thriving or atrophying.
Duplicate detection
Cosine similarity scans pair near-identical nodes automatically. One tap to review and merge — the healing journal records every decision.
Healing journal
Immutable audit log of every consolidation decision. Who merged what, when, and why — forever.
Contradiction detection
Flags nodes asserting conflicting facts about the same entity. Surface the conflict; you decide which version survives.
Auto-relinking
After a correction, broken references are healed forward through the graph. Structure is preserved even as content evolves.
Temporal decay
Nodes not reinforced by recall gradually decrease in weight. Your cortex stays fresh — stale facts fade, active ones strengthen.
Layered memory architecture
.gai Your attested canonical memory. Only you change this.
Opt-in · Local Ollama · Every inference stays on your device
“When you want the machine to think, not just search — without sending your memory to anyone else’s model.”
Recall enrichment
The local LLM rewrites your query at search time — adding synonyms, bridging languages — before it hits the index. Your search intent, amplified.
Correction parsing
"The deadline moved to March" becomes a structured diff against your existing nodes. Plain language in, precise graph edit out.
GNN edge prediction
A graph neural network proposes connections between nodes your explicit notes never linked. Surfaced as predictions; you promote or discard.
GLL assertion synthesis
The LLM reads co-recalled nodes and proposes atomic facts that follow from them. Always labelled as inference — never silently written as truth.
Insights & predictions
Background analysis of patterns, gaps, and risks across your full cortex. Strategic memory surfaced without you having to ask.
Layered memory architecture
.gnn Neural network edge predictions. Reviewed, not automatic.
.gll LLM assertions and inferred edges. Always labelled, never merged without approval.
The LLM never writes to your canonical memory. It advises. You decide. The architecture makes that guarantee structural, not procedural.
How it compares
| Feature | MemoryStudio | Cloud memory services |
|---|---|---|
| Data location | Your disk, encrypted | Their servers |
| Recall method | Deterministic subgraph | Probabilistic lookup |
| Correction flow | Diff → approve → apply | Opaque rewrite |
| Offline capable | ✅ Always | ❌ Requires internet |
| Audit trail | Op-log on your device | None / inaccessible |
| LLM | Local Ollama (optional) | Their cloud model |
| Per-query cost | $0 | API fees accumulate |
| Subgraph visibility | Full — nodes, scores, edges | Hidden from you |
Memory Kits
MemoryStudio is most powerful when your engrams already have structure. Graphnosis Memory Kits are pre-built engram templates for specific jobs — researcher, engineer, clinician, executive — with the right sensitivity tiers, source scaffolding, and node schema pre-configured.
Load a kit, open MemoryStudio, and your first recall will already have something worth exploring. The Deterministic Consolidation engine and GNN have richer material to work with from day one.
Download Graphnosis, unlock your cortex, and open the MemoryStudio tab. The Studio subscription unlocks the full feature set — recall, remember, edit, GNN, and both background engines. Your data stays local from the first keystroke.
Free tier available · Studio subscription required for advanced features · macOS 13+ · Windows 10+