Why uniko¶

uniko is the embedded memory layer for AI agents. It links into your process like SQLite, costs $0 to ingest, and reasons over what it stores. This page covers the problem it solves, the category it defines, and how it compares.

Agents are stateless. Memory is the missing layer.¶

An agent can pull text snippets from a vector store, but it cannot track who said what across sessions, notice when a fact changes, learn a procedure from repeated experience, or explain why it believes something. Those are different jobs, and a flat embedding index does none of them.

A bolt-on stack charges you three times¶

The usual answer stitches together two to four systems:

graph LR
    A[Agent] --> V[Vector store<br/>Qdrant / Pinecone]
    A --> G[Graph DB<br/>Neo4j / FalkorDB]
    A --> R[Rules / inference<br/>LLM calls]
    V -.consistency glue.-> G
    G -.consistency glue.-> R
    A --> L[LLM on every write<br/>extraction]

• Infrastructure cost. Each backend is a service to deploy, secure, back up, and keep in sync. The graph and the vectors drift apart, and reconciling them is your problem.
• Ingest cost. Per-message LLM extraction is money and latency on the write path, and it makes ingest network-bound and offline-hostile.
• Inference cost, paid every time. When reasoning lives in an LLM at query time, you pay for the same derivation on every question.

uniko removes all three by design.

One typed knowledge graph, in your process¶

Memory is a single typed knowledge graph organized around communication, living inside one embedded database. Messages are the atomic unit; entities, observations, facts, procedures, and topics derive from them with full provenance. The graph, the vectors, the full-text index, and the Locy logic engine are the same database — uni-db — running in your process.

Every advantage follows from one decision¶

One graph, in-process, compiled at ingest. Each differentiator below is a structural consequence of that choice.

How uniko compares¶

uniko is the only embedded, zero-infrastructure memory system in its class, and it backs that with the numbers that gate a production deployment: $0 ingest cost, 33–76× faster ingest than graph backends, and the fastest Q&A latency of any system measured (4.04s). The tables below are head-to-head, every figure quoted from the source reports.

Where everyone sits¶

uniko scores 81.2% on the full LoCoMo10 (1,986 questions), with retrieval hit 85.6%, F1 0.321, and total LLM cost of $3.55 (gemini-3.1, 2026-05-26, using Mem0's verbatim judge prompt) — delivered from one embedded graph instead of a stack of external services.

Cost and latency vs the KTH dmas-memory baseline¶

uniko's strongest measured ground is ingest throughput / cost and end-to-end query latency. The figures come from the KTH dmas-memory comparison (Wolff & Bennati, KTH, arXiv:2601.07978), measured 2026-06-14, unconstrained network mode, over the same LoCoMo10 question set.

Who uniko is for¶

uniko is a Rust library that embeds cognitive memory into your agent's process. It fits when:

• You want zero operational footprint. No external graph DB, no vector store, no managed service. The whole memory system — database, NLP cascade, embeddings, optional reranker — runs in-process on consumer hardware.
• Ingest cost and offline capability matter. Local extraction means predictable, $0-token ingest with no per-message network dependency.
• Conversation and provenance are central. You track who said what across many sessions, attribute statements to speakers, and explain why a fact is believed — not just retrieve nearby text.
• You organize memory around goals. Working memory assembled by traversing Goal → Task → Session is a first-class concept, not something you reconstruct yourself.
• You value graph-native reasoning. Facts, procedures, and topics are compiled at ingest and queried directly, with Locy logic available inside the database.

uniko is built for what no other shipping system offers: an embedded, conversation-native, goal-oriented memory graph with formal reasoning, at zero ingest cost. If your product depends on remembering people, facts, and procedures across sessions — in your own process — this is the layer for it.

Next steps¶