Open-core memory infrastructure for AI

Your agent forgot.
Now it doesn't.

Spectral memory engine, Rust-powered RAG, native knowledge graphs, 3D visualization. Bring your own LLM, your own database, your own data.

Apache-2.0 engines · PostgreSQL-backed · Self-host or managed

Self-host the whole stack
$ docker compose -f docker-compose.bundle.yml up

You shouldn't ship your agent's long-term memory through a closed vendor.

The memory your agent accumulates is your IP — user intents, factual context, retrieval embeddings, audit trails. Closed memory backends mean your data lives on someone else's servers, indexed by their schema, queryable on their terms. Kumbukumbu engines are Apache-2.0. Self-host the whole stack, or use the managed SaaS — same engine in both.

Four engines, one stack.

Each engine ships as its own package. Run them together via the bundle, or pick the one you need.

ASMIS

Memory with shape

A fact, a how-to, a recent event, an uncertain hunch — they're different shapes of memory. Kumbukumbu retrieves the right shape, not just the closest match. Beyond vector similarity.

pip install kumbukumbu-asmis

RAG

Rust-powered retrieval

Four Rust crates: core, server, CLI, Python bindings. Benchmarks for chunking, embeddings, vector search, and cache — you measure, not trust marketing numbers.

pip install kumbukumbu-rag

Knowledge graphs

Built-in, not bolted-on

core.knowledge ships graph + vector together. Unified KMS layer — query memory by similarity AND by relationship.

from kumbukumbu_asmis.core.knowledge

Viz

3D memory visualization

Interactive 3D views of memory graphs, goal hierarchies, agent networks, temporal patterns. Browser-side, npm package.

npm i @theaistep/kumbukumbu-viz

Plus kumbukumbu-sidekick: an autonomous optimization service using LangGraph workflows to keep your memory layer healthy without manual tuning.

No vendor lock-in. By design.

Self-hosted means you keep the data and the bill. Kumbukumbu engines provide the memory layer; everything around it stays yours.

Bring your own LLM

OpenAI, Anthropic, Google, Mistral, or fully local. The engine is provider-agnostic.

Bring your own database

PostgreSQL is the supported backend. Same SQL connection your app already uses.

Apache-2.0 engines

ASMIS, RAG, Viz are all Apache-2.0. Read the code, fork it, audit it.

No phone-home

Self-hosted instances never call back. Telemetry is opt-in and clearly named.

Six SDKs. Two engines × three languages.

Thin clients for ASMIS and RAG, in the language your stack already speaks. No bundled HTTP client, no opinionated runtime.

ASMIS SDKs

pip install kumbukumbu-asmis-sdk-py cargo add kumbukumbu-asmis-sdk npm i @theaistep/kumbukumbu-asmis-sdk

RAG SDKs

pip install kumbukumbu-rag-sdk-py cargo add kumbukumbu-rag-sdk npm i @theaistep/kumbukumbu-rag-sdk

Pricing.

Self-host the engines for free, or use the managed SaaS. Same engines in both.

Prism

Try it

For evaluation

0 €
  • 100 MB memory
  • 10 000 embeddings / month
  • 1 app, 1 tenant
  • Kumbukumbu branding
Start free

Alchemy

Indie

First production app

29 €

/month

  • 5 GB memory
  • 500 000 embeddings / month
  • Fernet encryption at rest
  • 5 apps + basic audit logs
Start trial
Most picked

Harmony

Pro

Audit-ready memory

99 €

/month

  • 50 GB memory
  • 5 M embeddings / month
  • Premium memory types
  • Reranking + hybrid search
  • Federated memory (opt-in)
Start trial
Sovereign

Sovereignty

Enterprise

Your domain, your region

From 800 €

/month, contact for scope

  • Multi-region replication
  • White-label (custom domain)
  • SOC2 attestation
  • On-premise option
Contact us

Or run the engines yourself — Apache-2.0, no usage limits, no phone-home, your PostgreSQL.

Memory types, three tiers deep.

Memory isn't one kind of thing. A user fact, a procedural skill, a long document, a short conversation — these need different lifecycles. Kumbukumbu ships a three-tier registry: core types in the Rust engine, platform types added at startup, and your own types registered per-app.

Tier 1 · Core

Built into the engine

Seven types baked into the Rust engine, covering the canonical shapes:

document_part conversation fact context user_memory computed_statistic

Plus platform types (system, batch, document) added by ASMIS at startup.

Tier 2 · Your own

Register per-app types

Devs register custom types, scoped by tenant. Two tenants can register the same type_id with different schemas — no collision.

registry.register_app_type( "quiz_answer", "Quiz Answer", "User responses", tenant_id="acme", )

Tier 3 · Premium

Audit-shaped (Harmony+)

Three managed-cloud types with extra guarantees, for regulated workloads:

regulatory_evidenceaudit-ready compliance artifacts
audit_trailimmutable, tamper-evident write log
signed_intentintent records co-signed with a Sawabona key

Shipping AI under audit?

Kumbukumbu is the memory layer of the LLM Compliance Bundle — paired with build-time analysis (Jagora), runtime sandboxing (Zelo), and a signed evidence chain (Ushahidi).

See the bundle →

Also from theAIstep — the rest of the stack: