Every AI session today starts from zero. Your agents reconstruct your org's decisions, context, and institutional knowledge from scratch — burning tokens and losing precision on work that was already done.
Zdravo closes that gap with a memory layer built on AI. We embed, summarize, classify, and retrieve knowledge automatically, then serve it to any model through a governance gate. The result: agents that remember, reasoning that compounds, and token costs that fall as memory grows.
Conversations from ChatGPT, Claude, Gemini, Cursor and 10+ platforms are captured in one click via extension, bookmarklet, MCP, or CLI.
Every memory is converted into a 768-dimensional vector with nomic-embed-text (self-hosted Ollama), with OpenAI and Gemini as fallbacks. A strict dimension contract guarantees every vector is comparable.
A background worker calls an LLM to generate a concise summary, extract 3–5 tags, and classify the memory type (episodic, semantic, procedural, working, emotional). Enrichment is async so capture stays instant.
Vectors land in an HNSW-indexed vector store; structured metadata, provenance, and the encrypted payload live in Postgres behind row-level security. Provenance records which model and node produced each embedding.
Semantic search embeds the query and returns the most similar memories by cosine similarity — meaning, not keywords. Long memories are chunked and mean-pooled so nothing is lost.
ContextOS evaluates every retrieval against consent, residency, and quota policy before it is served. Nothing reaches an agent without passing the governance gate.
Memories are exposed to agents through a production-hardened MCP server and REST API, so Claude, Cursor, or your own code recall org memory in a single config line.
We are intentionally not locked to one vendor. Inference is local-first and every hosted call can be swapped for your own keys.
nomic-embed-text
Self-hosted Ollama · 768-dim
Fallback: OpenAI text-embedding-3-small · Gemini text-embedding-004
gpt-4o-mini
1–2 sentence auto-summaries
Fallback: Any OpenAI-compatible or local model via provider routing
LLM memory typing
Episodic · Semantic · Procedural · Working · Emotional
Fallback: Deterministic heuristics if the model is unavailable
Local-first Ollama
Run your own models, zero per-token cost
Fallback: Bring Your Own Key (BYOK) for any hosted provider
Agents and humans find the right memory by meaning. Embeddings turn questions and documents into the same vector space, so relevance beats exact-match keyword search.
Every saved memory is summarized, tagged, and classified by an LLM in the background — the knowledge base organizes itself as it grows.
When a teammate leaves, their institutional knowledge is transferred with original-author provenance. The org's memory survives offboarding instead of walking out the door.
A governance layer enforces consent, data-residency, and quota policy on every retrieval. Agents only receive memory they are authorized to see.
A production MCP server lets Claude, Cursor, and custom agents read and write org memory through one standard protocol.
Enterprise onboarding supports client-side encryption so content is encrypted before upload — the AI works on encrypted references, not raw content.
768
Dimensions per embedding, enforced by contract
10+
AI platforms captured into one memory layer
100%
Retrievals gated by ContextOS governance
0
Customer messages used to train models
We do not use customer content to train any model. Memory stays in your database; retrieval is the product, not the training set.
PII detection runs on every ingestion. Sensitive fields are flagged and handled under strict access and retention controls.
Every memory, retrieval, and agent decision is logged in a hash-chained audit trail. Show an auditor exactly what the AI knew and when.
RBAC, IP allowlists, and dual-admin approvals keep people in control of who and what the AI can access.
Bring your own models and keys, or self-host Ollama inference. When the next frontier model ships, your memory stays put.
GDPR-aligned deletion removes memories and their vectors from the vector store and Postgres on request — no shadow copies.
Today Zdravo is model-agnostic under the hood. Next, we're putting that power in your hands — pick any model at runtime, the way Kilo Code lets you switch providers in one click.
Alibaba — flagship open-weight models
Gemini family — multimodal, long context
Inference at wafer-scale speed
Reasoning-strong open models
Efficient European open-weight models
GPT-4o family — already supported
Claude family — already supported
Run any model locally, zero per-token cost