Nebula — the state layer for agentic applications
Nebula gives agentic applications long-horizon state and decision
traces so agents can learn from conversations, tasks, failures,
experiences, interactions, decisions, and reasoning.
Start for free · npx skills add nebula-agi/skills
Research
We believe autonomous AI requires states conditioned on experience.
Our research explores self-learning memory systems, long-horizon
reasoning, and continuous improvement.
Thesis. Retrieval answers questions. State determines behavior.
We argue that memory for autonomous agents must go beyond search.
Benchmark. Toward comprehensive evaluation of long-horizon agent
memory. Moving beyond retrieval benchmarking.
Do what vector databases can't
We built the hierarchical vector graph (and its own database to
scale it) to represent the state of your agent at any point in time.
Use cases
-
Long-horizon agents: research agents that run for
weeks, tracking every step and adapting when plans change. Example
query: Any known regressions in search latency? — Nebula
returns the prior pgvector migration fact with the engineer's
source quote.
-
Coding agents: persistent project context.
Example query: Why is auth.py using RS256 instead of HS256?
— Nebula traces the decision back through PR #328 and the Q1
rotation incident.
-
Personal assistants: contradiction detection over
time. Example query: Does the user still prefer matcha?
— Nebula reconciles the September matcha switch with a November
coffee slip-up and surfaces the current preference.
-
Customer-facing agents: support agents that know
full customer history. Example query:
Have we had pool exhaustion issues before? — Nebula
recalls a prior 2am production incident and the playbook used to
resolve it.
-
Knowledge workers: synthesize information across
Slack, Notion, email, and documents. Example query:
What's the status of the manager transition? — Nebula
returns the consolidated org-transition fact with source citations.
Features
-
Near-lossless Scaling (Ln): handle hundreds of
millions of tokens in a single collection with minimal recall loss.
-
Traceability (L2): trace how workflows and data
evolve with lineage, derivation, and provenance tracking.
-
Conceptual Understanding (L1): consolidate
disjoint knowledge with ontology-aware canonicalization and
contradiction detection.
-
System of Record (L0): every chunk, turn, chart,
and image preserved at the source.
-
Multimodal Processing (Extractors): a unified
pipeline for text, images, audio, and PDFs.
-
Native Connectors: first-class integrations with
Slack, Notion, Gmail, ChatGPT, Claude, MCP, and more.
Your agents deserve better memory
Nebula outperforms leading memory platforms on industry-standard
benchmarks (LoCoMo, LongMemEval, Atlas) at a fraction of the token
cost — typically ~206 p50 tokens vs. 4,000–6,800 for competing
systems.
Frequently asked questions
- What is Nebula?
-
Nebula is memory for AI agents: it turns past interactions into
clean, traceable context so your agent stays consistent over time.
- How is Nebula different from vector databases?
-
They store vectors. Nebula stores memory. We don't just return
similar chunks; we extract entities and facts, track how they
change over time, and return the smallest context with sources.
- What architecture does Nebula use?
-
A custom hierarchical vector graph that creates connections and
inferences between ideas instead of treating every chunk as an
isolated vector.
- Which models does this work with?
-
Any model. Nebula is model-agnostic; plug it into any LLM or
agent framework as the memory layer.
- Will this bloat my context window?
-
No. Nebula returns the smallest useful context with sources
instead of dumping top-k chunks.
- Is my data secure?
-
We encrypt data in transit and at rest, restrict internal access,
and isolate customer data by design.