Concept-graph indexing
Decompose any artifact into provenance-tagged concept nodes and link them by semantic + structural similarity.
- 2–7 nodes per artifact
- Provenance preserved
- Real-time ingest
Concept-graph indexing · semantic recall
A knowledge graph that thinks with you.
Mneme-Net is the persistent memory layer for AI products. We index your concepts as a graph, surface them with sub-millisecond recall, and pulse the parts of memory your model is actually using — so you can see the shape of what your product knows.
- 14msp99 recall over 1.2M nodes
- 3 SDKsPython · TypeScript · Rust
- SOC 2Type II · self-host or cloud
02 How it remembers
Concepts in,
graph out.
Every artifact you feed Mneme — a doc, a chat turn, a tool call, a customer email — is decomposed into concept nodes. We embed each node, link it to the concepts it touches, and write the edges to a graph store you control.
No more flat vector blobs. Memory becomes a place you can walk.
-
i
Decompose
Source artifact → 2–7 concept candidates with provenance.
-
ii
Embed & link
Each concept embedded with the model of your choice; edges added by semantic + structural similarity.
-
iii
Write to graph
Persisted to your Postgres + pgvector, your Neo4j, or our managed store. Your data, your row-level policy.
03 How it retrieves
Edges that
travel light.
On every recall, we walk the graph from your prompt's nearest concepts outward, scoring edges by semantic strength and your own usage signal. The matching cluster lights up in the rendering — those are the concepts your model is about to think with.
You can debug an entire memory query by watching one frame.
14ms
p99 over 1.2M nodes
2.4ms
median 2-hop recall
99.97%
retrieval availability
$0.012
per 1,000 recalls
04 Capabilities
Everything your AI
needs to remember.
A complete memory layer. Drop-in SDKs for the languages you ship in, observability for the team that operates the product, and primitives the founders can compose on day one.
Semantic recall
Walk the graph from your prompt's nearest concepts outward. Sub-millisecond, deterministic, debuggable.
- k + hop + budget knobs
- Per-edge usage signal
- Frame-by-frame debugger
Graph observability
See which concepts your product is actually using, which clusters are dead, which edges fire on every chat.
- Cluster heatmap
- Edge fire-rate
- Drift alerts
Bring your own store
Postgres + pgvector, Neo4j, or our managed store. Your data, your row-level policy, your region.
- 3 store backends
- RLS & SOC 2
- EU + US regions
SDKs for shipping teams
Python, TypeScript and Rust — each with the full graph + recall surface and zero ceremony.
- Strict types
- Streaming recalls
- Edge runtime safe
Memory provenance
Every recalled concept carries the artifact it came from. Audit, redact, and forget at the source.
- Source-of-truth links
- GDPR-grade forget
- Full audit log
05 Pricing
Three ways
to remember.
Start free in the Sandbox. Move to Workshop when your product ships. Foundry when memory becomes the moat.
Feature
i
Sandbox
$0/ mo
Free, no card
Popular
ii
Workshop
$249/ mo
Default for teams
iii
Foundry
Custom
Self-host & SLA
Concept nodes
10,000
1.2M
Unlimited
Recalls / month
50,000
5,000,000
Unlimited
Graph store
Managed
Postgres · Neo4j · managed
Self-host or dedicated cloud
Observability suite
—
✓
✓
SOC 2 Type II + DPA
—
—
✓
Named graph architect
—
—
✓
Support channel
Community
Email · Slack
On-call · 99.99% SLA
All tiers include the full SDK surface, graph debugger, and provenance. Annual billing — 20% off.
06 Contact
Memory
that ships with you.
Drop a line and we'll spin up a graph from your existing artifacts in under a week. No procurement maze.