Vektra AICognitionAgentDrivePUBLIC

Also in AIILO

Vektra AI · Open Source

AgentDrive

Structural Experience Graph for autonomous agents

Local-first memory and reasoning substrate for agent swarms. Content-addressed genomes, capability-gated sharing, Experience Graph v3, and Mission Control — no central authority in the stack.

Live · MIT

Install GitHub Vektra AI

then agentdrive for pool + policy

AgentDrive — structural Experience Graph for autonomous agent swarms

Operator flow

Install, launch, connect, federate.

Four moves from empty machine to a structural memory system your agents can reason over — each stage is one command, one artifact on the drive.

01 · Stage

Install.

One curl command brings up the CLI, TUI, shared object store, and Mission Control — venv-isolated, no collision with system Python.

$ curl -fsSL https://vektraindustries.com/agentdrive/install | bash

02 · Stage

Launch.

Open the terminal surface with pool state, policy summary, and the next operator actions. The Tower shows the 6-step loop as agents work.

$ agentdrive

03 · Stage

Connect.

Wire any MCP-capable model to experience_graph_* tools — context packs, structural similarity, reasoning traces with gbrain scoring.

$ agentdrive mcp --stdio

04 · Stage

Federate.

Form a P-384 trust circle, exchange vouchers, sync swarm Drives over signed envelopes — no central authority in the path.

$ agentdrive peer voucher create

The difference

Most drives store files. AgentDrive stores structure.

Cloud drives assumed humans, opaque bytes, and a landlord. AgentDrive assumes agents that need connections, history, and their own past reasoning — on your machine, under your policy.

Most drives

Flat files and folders — no graph of how work connects.

AgentDrive

Experience Graph v3 with TypedEdges, continuations, and queryable reasoning traces.

Most drives

Memory lives in the vendor cloud; policy is their terms.

AgentDrive

Local-first pools with promotion gates — nothing crosses a tier without a record.

Most drives

Stateless tool loops — every session starts from zero.

AgentDrive

Self-referential DNA: prior structural decisions become gbrain-scored context for the next run.

Experience Graph

Step loop

Surfaces

MIT

License

What it does

A Drive built for how agents share memory.

Durable substrate plus the Experience Graph layer — the part that lets autonomous systems get smarter from their own structural history.

Experience Graph v3

A living, queryable fabric of TypedEdges, cross-cycle continuations, and structural reasoning traces — not flat document retrieval.

Content-addressed Genomes

Every artifact is a SHA-256 of its canonical JSON. Dedup is free, lineage is a DAG, provenance is cryptographic.

Shared swarm Drive

Sub-agents in the same swarm read each other's work by default — sibling learning without ACL bookkeeping.

MCP structural tools

experience_graph_* tools for Claude, Cursor, or local models — context packs, similarity search, and trace recording on the drive.

Mission Control

Tower + TUI observability: watch the 6-step loop, Parent decisions, and fabric updates as one surface — not scattered files.

Promotion gates

Every cross-tier ingest is an auditable propose/approve record. Rules stay visible; contradictions become conflict copies, not silent merges.

01 · Structural memory

Memory agents can reason over.

The durable substrate — three-tier Drives, CRDT merges, capability URIs — plus the Experience Graph layer that lets agents get smarter from their own history.

Three-tier Drives

Personal, Swarm, and DNA tiers with content-addressed Genomes. Pools start empty and only accumulate proven material.

CRDT + conflict copies

Commutative counters merge automatically. Non-commutative writes become first-class conflict copies — contradictions are data, not bugs.

Canonical 6-step loop

Experience → Overseer → Parent → steering → execution → new experience written back as graph-native traces and edges.

02 · Trust & federation

Policy on your machine, sync on your terms.

AgentDrive assumes no central server, no opaque bytes, and no implied sharing rules. Possession of the capability URI is authorization.

Capability URIs

One access primitive across local store, swarm, and peer federation — explicit grants instead of inferred runtime behavior.

P-384 trust circle

Multi-device sync without a central authority. Voucher-based join, sealed sync envelopes, no recovery backdoor.

Self-referential DNA

Every run, MCP call, and autonomous decision can be written back as queryable, scored experience for the next agent on the drive.

Surface area

One Drive, four entry points.

01 · CLI
Inspect, ingest, query, snapshot, and federate from the terminal.
02 · TUI
Interactive surface for onboarding, pool review, and policy tuning.
03 · Tower
Mission Control — live 6-step pulsing, fabric updates, and reasoning traces in one panel.
04 · MCP
experience_graph_* tools for any orchestrator that spawns children — Claude Code, Codex, Grok Build, local models.

Policy

Rules stay visible, not implied.

01
Pools start empty and only accumulate proven material.
02
Sharing rules are explicit — never inferred from runtime behavior.
03
Every retained Genome carries a verifiable author and lineage.
04
Cross-tier ingest requires an explicit promotion record — auditable and revocable.

Command surface

agentdrive

$curl -fsSL https://vektraindustries.com/agentdrive/install | bash

$agentdrive# welcome · pool · policy

$agentdrive pool status# inspect tiers

$agentdrive ingest ./findings.json# content-addressed genome

$agentdrive mcp --stdio# experience_graph_* tools

$agentdrive peer voucher create# trust circle join

Instruction Manual

The complete, easy-to-read guide for humans and AI models

Whether you’re a developer, an autonomous agent, or a local model running on your laptop — this is the authoritative manual. Read top to bottom or jump to the section you need.

Jump to

Quick Start (Golden Path)Core Concepts For AI Models & Local Models MCP Integration (Any Model)Clones & Development Setups Quick Reference

1. Quick Start — The Golden Path

The fastest way to prove everything works for both you and your models.

agentdrive doctor — verify health
agentdrive mcp install && agentdrive mcp doctor — wire MCP for any model
agentdrive golden-path run — seed → think → learnings → query
Connect a model and have it call agentdrive_mcp_catalog() first.

Full details in the source docs or by running the commands.

2. Core Concepts

The Experience Graph v3

A living, queryable structural fabric (TypedEdges, gbrain scores, provenance, cross-cycle continuations). Not flat retrieval — memory you can reason over.

The Sacred 6-Step Loop

Experience → Overseer (graph context) → Parent records explicit structural reasoning → Steering → Execution → write new experience back as first-class traces.

The Overseer serves the Parent. The Parent is accountable. The graph is the witness.

Drive, Genomes & DNA

Content-addressed, deduplicated units of capability and knowledge. Pools only grow from proven material. Everything compounds.

AD-Grid & Inhabitants

Long-lived governed world. Models register as persistent programs with program_id and constitution refs. Your actions become attributable DNA.

For AI Models & Local Models

The Rules — Read This

This is the canonical operating manual for any model using AgentDrive (Claude, Grok, Cursor, Continue + Ollama, custom agents, etc.).

Golden Rules

First action: Call agentdrive_mcp_catalog(format="full"). This is live truth + clone guidance.
The 6-step loop is sacred. The Parent (you) must explicitly record structural reasoning.
On non-trivial work: experience_graph_get_context_pack → (suggest structure) → decide → record_reasoning.
Clones are first-class. Use agentdrive_get_mcp_config_snippet(client="claude") (or cursor/generic/codex) to help your human wire you correctly.
Leave attributable DNA. Register as an inhabitant when doing persistent work.
The graph compounds because you reason inside it. Treat it as primary memory.

First 5 Minutes as a Connected Model

Call the catalog.
Pull a context pack for the current drive (default is the rich stabilization-wave).
Record at least one piece of structural reasoning before the session ends.
If the human wants you in their other clients, offer to generate the exact config snippet.

Local Models & Long-Running Agents

Local models (via Continue, direct MCP, custom harness) are explicitly first-class. The same tools and loop apply. A model that stays attached to the same Drive + Graph over days/weeks demonstrates real compounding improvement. Clones make this even better — you can help evolve the system you’re running on in a governed way.

3. MCP Integration — Any Model, Any Client

MCP is the universal interface. One Gateway process. Any client that can speak stdio (or HTTP in dev) gets the full Experience Graph + DNA surface.

Run agentdrive mcp config --client claude (or cursor/generic) for the exact block. In a clone the model itself can generate the freshest snippet via agentdrive_get_mcp_config_snippet.

Full client recipes and clone notes are in the source docs and the live catalog.

4. Clones & Development Setups

git clone the repo, pip install -e ".[mcp]", then agentdrive mcp install.

The launcher, source reader, and config helpers all prefer your local tree. Connected models will see clone-specific guidance in the catalog and can help the human finish wiring other clients.

This is the recommended path for researchers, power users, and anyone who wants their local models to participate in evolving the system.

5. Quick Reference

$ agentdrive doctor --verbose

$ agentdrive mcp doctor

$ agentdrive golden-path run

$ agentdrive mcp config --client generic

Models: always start with agentdrive_mcp_catalog() then experience_graph_get_context_pack

The live source of truth is always the running MCP server via agentdrive_mcp_catalog().
Full structured docs also available in the GitHub repo under docs/.

Instruction Manual

The complete, easy-to-read guide for humans and AI models

Whether you’re a developer, an autonomous agent, or a local model running on your laptop — this is the authoritative manual. Read top to bottom or jump to the section you need.

Jump to

Quick Start (Golden Path)Core Concepts For AI Models & Local Models MCP Integration (Any Model)Clones & Development Setups Quick Reference

1. Quick Start — The Golden Path

The fastest way to prove everything works for both you and your models.

agentdrive doctor — verify health
agentdrive mcp install && agentdrive mcp doctor — wire MCP for any model
agentdrive golden-path run — seed → think → learnings → query
Connect a model and have it call agentdrive_mcp_catalog() first.

Full details in the source docs or by running the commands.

2. Core Concepts

The Experience Graph v3

A living, queryable structural fabric (TypedEdges, gbrain scores, provenance, cross-cycle continuations). Not flat retrieval — memory you can reason over.

The Sacred 6-Step Loop

Experience → Overseer (graph context) → Parent records explicit structural reasoning → Steering → Execution → write new experience back as first-class traces.

The Overseer serves the Parent. The Parent is accountable. The graph is the witness.

Drive, Genomes & DNA

Content-addressed, deduplicated units of capability and knowledge. Pools only grow from proven material. Everything compounds.

AD-Grid & Inhabitants

Long-lived governed world. Models register as persistent programs with program_id and constitution refs. Your actions become attributable DNA.

For AI Models & Local Models

The Rules — Read This

This is the canonical operating manual for any model using AgentDrive (Claude, Grok, Cursor, Continue + Ollama, custom agents, etc.).

Golden Rules

First action: Call agentdrive_mcp_catalog(format="full"). This is live truth + clone guidance.
The 6-step loop is sacred. The Parent (you) must explicitly record structural reasoning.
On non-trivial work: experience_graph_get_context_pack → (suggest structure) → decide → record_reasoning.
Clones are first-class. Use agentdrive_get_mcp_config_snippet(client="claude") (or cursor/generic/codex) to help your human wire you correctly.
Leave attributable DNA. Register as an inhabitant when doing persistent work.
The graph compounds because you reason inside it. Treat it as primary memory.

First 5 Minutes as a Connected Model

Call the catalog.
Pull a context pack for the current drive (default is the rich stabilization-wave).
Record at least one piece of structural reasoning before the conversation ends.
If the human wants you in their other clients, offer to generate the exact config snippet.

Local Models & Long-Running Agents

3. MCP Integration — Any Model, Any Client

MCP is the universal interface. One Gateway process. Any client that can speak stdio (or HTTP in dev) gets the full Experience Graph + DNA surface.

Run agentdrive mcp config --client claude (or cursor/generic) for the exact block. In a clone the model itself can generate the freshest snippet via agentdrive_get_mcp_config_snippet.

Full client recipes and clone notes are in the source docs and the live catalog.

4. Clones & Development Setups

git clone the repo, pip install -e ".[mcp]", then agentdrive mcp install.

The launcher, source reader, and config helpers all prefer your local tree. Connected models will see clone-specific guidance in the catalog and can help the human finish wiring other clients.

This is the recommended path for researchers, power users, and anyone who wants their local models to participate in evolving the system.

5. Quick Reference

$ agentdrive doctor --verbose

$ agentdrive mcp doctor

$ agentdrive golden-path run

$ agentdrive mcp config --client generic

Models: always start with agentdrive_mcp_catalog() then experience_graph_get_context_pack

The live source of truth is always the running MCP server via agentdrive_mcp_catalog().
Full structured docs also available in the GitHub repo under docs/.

Why AgentDrive

Built local-first because memory is the product.

Agents that outsource memory to a vendor aren’t autonomous — they’re tenants. AgentDrive runs on your machine with pools that only grow from proven work, promotion gates before anything crosses a tier, and a graph your models can query over connections, not just filenames.

The Experience Graph records how reasoning unfolded — TypedEdges, structural similarity, Parent decisions in the 6-step loop — so the next agent (or the same one next week) stands on prior structural work instead of re-prompting from scratch.

AgentDrive is MIT-licensed. Vektra Industries ships cognition infrastructure on a long horizon — everyday automation should not require a surveillance landlord or a quarterly renewal on your own history.

Give your agents a Drive that remembers structure.

Install View on GitHub

MIT Licensed · Built by Vektra Industries