Skip to content
VENKAI
Pilot

Research

Verified, experimental, planned, hypothesis.

Four separate columns, never mixed. A capability that exists in the repository is not thereby a demonstrated capability.

Quarantine

The older benchmarks are no longer published

This site published cost-reduction and continuity figures for a long time. An adversarial audit run on 10 August 2026 showed those measurements came from a harness that modelled the saving instead of measuring it, with several values hardcoded in the benchmark script itself.

They were withdrawn from the site rather than reworded. A real LLM cost measurement requires running real agents with and without Venkai, on the same work; that work hasn't been done. Until it is, Venkai puts forward no cost figure.

Status

The state of each piece

  • VERIFIED

    Persistence and handoff

    80/80 core tests, 19/19 handoff checks, including a handoff to an agent running a different model.

  • VERIFIED

    Ranked retrieval

    Similarity, recency, importance and frequency, covered by dedicated tests including a self-reinforcement guard.

  • VERIFIED

    Isolation, versioning, access

    Org filtering, snapshot on every write, JWT and API keys, GDPR export and deletion.

  • EXPERIMENTAL

    Security knowledge graph

    CWE/CAPEC/OWASP import and AST analysis; tests pass, but nothing has been validated on live data. Independent from the product core.

  • EXPERIMENTAL

    Semantic Modification Engine

    A product line distinct from this layer: minimal, AST-guided code edits. The engine runs and is test-guarded; its previously published figures have been withdrawn and are not replaced here. It is not what the continuity layer does — which is not the same as abandoned.

  • EXPERIMENTAL

    MCP server in the field

    Seven tools tested locally, never wired to a third-party MCP client.

  • PLANNED

    PostgreSQL and deployment

    Migrations written, Docker and Caddy present, neither executed nor validated.

  • PLANNED

    Semantic embeddings

    Ranking accepts a real embedding model; the default stays a hash with no generalisation.

  • PLANNED

    Feedback loop

    Positive and negative feedback is collected, but does not yet influence ranking.

  • LIMIT

    Governance layer

    It exists in the repository and is wired to neither the API, nor the SDK, nor MCP. It is therefore not a feature and is not presented as one.

Hypotheses

What we believe without having shown it

Stated as hypotheses because that's what they are. Nothing here should be read as a result.

  • State weighs less than history

    Passing typed objects should cost less than a conversation dump, at equal resumption quality. Not measured.

  • The real cost is undone work

    We assume lost context costs mostly in repeated work and reversed decisions, more than in tokens. Not measured.

  • Writing is the hard part

    We assume the adoption difficulty is getting agents to write at the right moment, not getting them to read. To be checked in a pilot.

Build a continuity layer for your agents.

Describe your multi-agent workflow and where context goes missing. We answer with a pilot scope, or with the reason Venkai isn't the right piece.

Describe your workflow