The vocabulary of agentic infrastructure.
Kamino builds agentic infrastructure — the operating environment where autonomous systems remember, adapt, and endure. These are the terms that vocabulary rests on: each defined in a sentence or two, and pointed back to where in the ecosystem it lives.
The persistent sense of self an AI agent accumulates over time — a voice, a perspective, and self-awareness that emerge as knowledge crystallises, so the agent carries a stable identity from one session to the next rather than resetting to zero. It is the top layer of Elephantasm's memory hierarchy and the framework's namesake capability.
Elephantasm's four-layer model of agent memory, ordered by abstraction and stability: raw Events are synthesised into Memories, Memories are distilled into canonical Knowledge, and accumulated Knowledge crystallises into Identity. Each layer is more stable and more abstract than the one beneath it.
Elephantasm's retrieval method. Rather than ranking memories by semantic similarity alone, it scores each one across four independent factors — importance, confidence, recency, and decay — and combines them with similarity into a single weighted composite. A critical memory from last week can therefore outrank a tangentially similar one from yesterday.
Elephantasm's autonomous memory-curation loop, inspired by sleep-based consolidation. Light Sleep runs fast, deterministic passes — updating decay scores, transitioning memory states, and detecting merge candidates by vector similarity — while Deep Sleep brings in an LLM to evaluate merges, review flagged memories, and split conflated concepts. Every action is logged with before-and-after snapshots, and nothing is truly destroyed.
A governance layer between AI agents and payment providers. Every transaction must carry a structured justification and pass policy checks before money moves, and each step is written to an immutable audit ledger. In the Kamino ecosystem this is C.R.E.A.M.
Turning raw logs from any provider and format into structured, canonical, token-efficient events — classified by meaning rather than by fragile regex patterns. In the Kamino ecosystem this is Lumber, which uses a local embedding model to file each log line into one of a fixed set of categories.
A database that stores high-dimensional embedding vectors and retrieves them by nearest-neighbour similarity search, such as cosine distance or approximate nearest-neighbour. It is a component that memory frameworks build on — Elephantasm, for example, uses PostgreSQL with the pgVector extension — not a substitute for one.
An autonomous software system that can observe, decide, and act on its own toward a goal, rather than only responding when it is prompted. Kamino's frameworks give agents the memory, identity, perception, and means to act that let them operate continuously.
The finite span of text a language model can consider at once. Context windows have grown past a million tokens, yet a model still resets to amnesia after each session — which is why persistent memory, not a larger window, is what gives an agent continuity.
Software that operates with continuity, identity, persistent memory, and the ability to improve through feedback — where its own outputs become the inputs of the next cycle — instead of finishing a task and vanishing. Building the operating environment for such systems is Kamino's single goal.