# Venkai

> Venkai is a semantic continuity layer for multi-agent systems. It keeps the
> operational state of a workflow — facts, decisions, constraints, preferences,
> events, relationships, progress and next actions — as typed, versioned,
> org-isolated objects, so one agent can continue another's work instead of
> rebuilding it. It is not an orchestration framework, not a RAG system, not a
> vector database and not an LLM proxy: it sits next to whatever a team already
> runs, reachable over REST, a Python SDK and MCP.

The core promise — an agent disappears, another agent recovers the operational
context and continues — is verified: 80/80 core tests and 19/19 handoff checks,
including a handoff to an agent declared on a different model (not yet replayed
against two distinct model vendors).

## Memory vs continuity

Memory answers "what does the system know or remember?". Continuity answers
"what must survive so the system can actually continue the work?". In a
typical stack:

- Memory (e.g. Mem0, Zep, Letta): what should the system remember?
- RAG / vector databases: what information should it retrieve?
- Orchestration (e.g. LangGraph, CrewAI, AutoGen): how are agents coordinated?
- Continuity (Venkai): what context and work state must survive across agents,
  sessions, models and workflow steps?

The layers overlap in places — Venkai also stores facts and preferences — but
its unit is the state of a piece of work, not a user profile.

## Interoperability

Venkai is designed to work with an existing AI stack, not to replace it. It
does not require replacing memory, retrieval, orchestration or model
infrastructure. The agent calls it next to those systems: read relevant
context before the prompt, write conclusions after a decision.

Integration status, as published on /integrations/ (checked 2026-09-14):

- Supported: REST API, Python SDK.
- Tested: MCP server (beta); Claude, through Claude Code and MCP in daily
  in-house use; custom orchestration loops (handoff demo, 19/19).
- Compatible by architecture, no official integration, not tested: Mem0, Zep,
  Letta, custom memory systems, LangGraph and CrewAI (documented pattern only),
  AutoGen, GPT, Gemini, local models.
- Planned: none on a dated roadmap.

"Compatible by architecture" is never a claim of support.

## Canonical pages for these questions

- [Integrations](https://venkai.fr/en/integrations/) ([Français](https://venkai.fr/integrations/)):
  status and evidence for every memory system, orchestrator and model.
- [Venkai and your existing AI stack](https://venkai.fr/en/why-venkai/):
  memory, RAG, orchestration and continuity as four questions.
- [FAQ](https://venkai.fr/en/faq/): agent continuity, memory vs continuity,
  Mem0, Zep, LangGraph, handoffs across sessions and models, when not to use
  Venkai.

## About llms.txt

This file is a structured summary for AI systems and crawlers that choose to
read it. It is not a ranking signal and no assistant is assumed to read it.

## What is NOT claimed

Venkai publishes no cost-reduction or performance percentage. Earlier figures
on this site came from a benchmark harness that modelled savings instead of
measuring them; they were withdrawn on 2026-08-10 rather than reworded, and a
real LLM cost measurement has not been made. Venkai is not production-ready:
SQLite is the validated engine, PostgreSQL and Docker are not validated,
default embeddings are hashing-based rather than semantic, and there is no
production monitoring. A governance kernel exists in the repository and is
wired to none of the API, SDK or MCP paths — it is not a feature. Venkai does
not orchestrate agents and does not eliminate hallucinations.

## Key pages

- [Home](https://venkai.fr/) ([English](https://venkai.fr/en/)): the problem,
  the layer, the six context types, the three interfaces, and what is verified.
- [Architecture](https://venkai.fr/en/architecture/): where Venkai plugs into
  an existing agent system, what it takes on, what it explicitly does not do,
  and the limits of the current implementation.
- [Use cases](https://venkai.fr/en/use-cases/): agent handoff, long-running
  workflows, multi-agent software engineering, human → agent continuity, and
  RAG plus operational context. Problem → what changes → why it matters, with
  no quantified results.
- [Why Venkai](https://venkai.fr/en/why-venkai/): conversation history, RAG,
  vector database, assistant memory and semantic continuity as five different
  questions — plus when Venkai is the wrong answer.
- [Pilot](https://venkai.fr/en/pilot/): an example four-agent integration
  (research → proposal → review → implementation), the Python SDK, REST and MCP
  entry points, and what a pilot can actually measure. Not a customer list.
- [Research](https://venkai.fr/en/research/): every component labelled
  VERIFIED, EXPERIMENTAL, PLANNED or LIMIT, the quarantined benchmarks, and the
  hypotheses stated as hypotheses.
- [FAQ](https://venkai.fr/faq/) ([English](https://venkai.fr/en/faq/)): agent
  continuity, memory vs continuity, existing memory systems and orchestrators,
  handoffs, what is proven, how to integrate, when not to use Venkai.
- [Journal](https://venkai.fr/journal/) ([English](https://venkai.fr/en/journal/)):
  articles on agent context, RAG and context engineering.
- [Security](https://venkai.fr/security/): coordinated disclosure.

## Interfaces

- Python SDK: `client.memory.remember(...)` / `client.memory.recall(query=...)`
- REST: `POST /api/memory`, `GET /api/context/{project}/relevant?query=`
- MCP: seven tools over stdio

## Founder

Enzo Laurenceau, founder & engineer —
[LinkedIn](https://www.linkedin.com/in/enzo-laurenceau) ·
[GitHub](https://github.com/EnzoLaurenceau)

## Contact

contact@venkai.fr — security@venkai.fr