Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.crosmos.dev/llms.txt

Use this file to discover all available pages before exploring further.

Every AI agent has the same problem. It listens, responds, and forgets. The next conversation starts from zero. Every time. Crosmos fixes that. It gives your agent a memory that grows with every interaction, so responses get better, not just bigger.

The problem with how agents remember today

Most tools chunk your documents, embed them in a vector database, and call it memory. That works for finding a paragraph about quarterly revenue. It falls apart the moment you need something like:
“What was I working on last summer?”Vector search finds text that mentions “summer.” It has no idea when the event happened or that a newer fact might supersede an old one.
These are not edge cases. They are the questions your users actually ask. And standard RAG cannot answer them.

How Crosmos is different

Crosmos does not flatten content into vector chunks. It builds a Monotonic Temporal Knowledge Graph: a living, time-aware graph of entities and relationships that only grows.
1

Ingest

You send content. Conversations, PDFs, plain text. Crosmos accepts all of it.
2

Extract and connect

Facts, entities, and relationships are automatically pulled from raw content and woven into the knowledge graph. No manual tagging. No prompt engineering.
3

Retrieve

When your agent asks a question, four independent search signals fire in parallel and return exactly the context it needs. Read more about how retrieval works.
That is the entire loop. Ingest, connect, retrieve.

Why it works

The graph only grows

New facts append. Old facts stay. Both “I work at Google” and “I joined Anthropic” coexist, timestamped and ranked by recency. History is always recoverable. A smart forgetting layer fades low-importance noise while keeping what matters accessible. The result: your agent builds a richer picture with every conversation.

Time is built in, not bolted on

Every fact carries two timestamps: when the event actually happened, and when the system learned it. Temporal queries are native. Ask “what changed since we last talked” and Crosmos uses the real event time, not the ingestion time.

Four signals, one answer

No single search method catches everything. Crosmos runs four in parallel and fuses them:
SignalWhat it catches
SemanticMeaning-level matches through vector similarity
KeywordExact names and terms that semantics miss
GraphContextually connected entities beyond text overlap
TemporalFacts inside the queried time window
Agreement across signals naturally boosts the best results. Each one covers the others’ blind spots.

More than memory, a reasoning layer

Dozens of conversations about programming languages, editor preferences, and tooling choices. Individually, they are scattered facts. Together, they tell you something bigger: this user strongly prefers low-level systems work with keyboard-centric tools and avoids GUI-heavy workflows. Crosmos clusters related memories into synthesized summaries in the background. It is not just storing facts. It is understanding patterns. Your agent does not just remember what was said. It reasons about what it all means. Learn more about how clusters work in the Clusters deep dive.

Structured, not stuffed

Raw text goes in. Structured knowledge comes out. Conversations, PDFs, and documents become entity-relationship graphs with confidence scores and provenance. The pipeline handles extraction automatically. See how memories are structured and stored.

Built for the tools you use

MCP server

Plug into any MCP-compatible client. Your agent gets memory automatically.

Connectors

Ingest from the tools you already use. No custom pipelines needed.

Multi-tenant

Isolated memory spaces per organization. Zero cross-contamination.

Any content type

PDFs, plain text, conversation transcripts. The pipeline handles them all.

Memories as the source of truth

Anthropic’s research on contextual retrieval makes the case clearly: agents need richer, more connected context to reason well. Raw chunks are not enough. Crosmos is built around that principle. Memories carry entity relationships, temporal grounding, and importance scores. Clusters synthesize patterns across those memories. Together, they form the best corpus an LLM could ask for: structured, connected, and ranked by what actually matters. When your agent retrieves context from Crosmos, it is not getting a bag of text fragments. It is getting a reasoning-ready knowledge base that improves with every conversation.

Next steps

Quickstart

Make your first API call in minutes.

How it works

Dive into the architecture behind the engine.