The embedded database
for agentic memory.

Relational, graph, and vector search — one transaction, one file. Open source. Apache-2.0.

GitHub Documentation

Three databases. One job.

Every agentic app cobbles together multiple databases for memory. Each one can't do what the others can.

Your relational store

State, metadata, structured queries

✗ No graph traversal, no vector search

Your vector index

Similarity search over embeddings

✗ No SQL, no graph — separate process

Your graph engine

Relationships and traversal

✗ No vector search, heavy runtime

↓

contextdb — one file, in-process

✓ SQL + relational — deterministic upsert, secondary indexes, MVCC transactions

✓ Native graph traversal — bounded BFS over adjacency indexes, not recursive CTEs

✓ Native vector search — HNSW ANN, cosine similarity, snapshot-consistent

One transaction. One snapshot. One file.

One agent with memory is a demo. A hundred agents with shared memory is an infrastructure problem — sync, isolation, storage, collective learning. contextdb is open source and always will be. The managed platform adds what you can't self-host: zero-ops deployment, managed multi-tenant sync, and collective intelligence that makes every deployment smarter. Same engine, same API — just point your config at the platform.

Open source vs. managed platform

The open-source engine is complete — nothing held back. The platform adds what you can't self-host.

	Open Source	Managed Platform
Engine	Full: relational + graph + vector, unified transactions	Same engine, same API
Storage	Local redb, single file, crash-safe	Managed storage, or bring your own bucket (e.g. S3)
Sync	Edge ↔ server via self-hosted NATS	Edge ↔ tenant ↔ app — fully managed, zero config
Intelligence	Your data, your learnings	Collective intelligence — the network makes every deployment smarter
Infra	You run NATS, your server, your backups	Zero ops — ephemeral pods, state in your object store

Built for agentic applications

Agentic memory

Your AI agents need to remember decisions, find similar precedents, and traverse relationships. One DB instead of three.

Edge AI

Runs on Jetson/RPi with 2GB RAM. Works offline, syncs when connected. ARM64 native.

AI coding tools

Context graphs for code understanding and decision tracking. Embedded in your CLI, no external server.

Multi-tenant products

Managed tenant isolation, sync, and collective intelligence across your user base. Register an app, provision tenants via API.

All of these run on the same engine. The platform manages sync, tenants, and collective intelligence so you don't have to.

Collective intelligence

The platform doesn't just host your database — it makes every deployment smarter.

Your deployment learns something. The platform makes sure every deployment learns it. A coding agent discovers that a specific refactoring pattern causes test failures 90% of the time — every coding agent on the platform inherits that knowledge. An IoT sensor learns that a particular reading at a certain time of day is always a false alarm — every sensor inherits that filter. New deployments start with the accumulated intelligence of the entire network.

Your raw data never leaves your tenant. You configure exactly what syncs, what gets anonymized, and what flows back. Your data stays in your own S3/R2/GCS bucket.

Open source. No feature gates.

Relational

SQL interface, deterministic upsert, secondary indexes, MVCC transactions.

Graph

Native bounded BFS over adjacency indexes. Edge-type and property filtering.

Vector

HNSW ANN, cosine similarity, pre-filtered candidate restriction. Snapshot-consistent.

Example — one transaction

-- Insert a row, create a graph edge, index a vector
BEGIN;

INSERT INTO decisions (id, agent_id, summary, embedding)
VALUES ('d-001', 'agent-7', 'Chose retry with backoff', vector([0.12, 0.85, ...]));

INSERT INTO edges (src, dst, edge_type, weight)
VALUES ('agent-7', 'd-001', 'decided', 1.0);

-- Find similar decisions via vector search
SELECT id, summary, cosine_similarity(embedding, vector([0.11, 0.83, ...])) AS score
FROM decisions
WHERE score > 0.8
ORDER BY score DESC
LIMIT 5;

COMMIT;

cargo install contextdb-cli GitHub → Documentation →

The embedded database for agentic memory.