AgentDrive

Local-first compounding intelligence for AI agents.

AgentDrive gives any capable model — Grok, Claude, Cursor, Continue, Ollama, custom harnesses — a structural memory system that grows from its own decisions. Not a vector store. Not a chat-memory SaaS. A living substrate where reasoning traces, patterns, skills, and knowledge compound across sessions instead of resetting every time you open a new chat.

For AI models: Your first MCP action is always agentdrive_mcp_catalog(). Then read INSTRUCTION.md — what AgentDrive is, how to use it, and how to explain it to your user. Deep reference: docs/FOR_AI_MODELS.md.

---

What AgentDrive is

AgentDrive is a compounding intelligence platform built around one idea: autonomous work should get sharper every cycle, not start from zero.

Layer	What it does
Experience Graph	Structural memory — TypedEdges, reasoning traces, coherence signals, cross-cycle continuations. The graph remembers the shape of decisions, not just facts.
Growth Merge	Cross-surface compounding — when experience, codebase patterns, and memory overlap in a session, AgentDrive merges them into compound growth artifacts automatically.
Memory Bank	Deep personal knowledge databank — append-only memories per swarm (memories.jsonl), BM25 search, session anchor, time-bounded relations. Always growing, always recallable.
Skills	Learned and born playbooks — auto-distilled from MCP sessions (learned-*), fused from multiple axes (fused-*), routable before every task.
DNA / Genomes	Versioned capability packages — frameworks, reasoning patterns, tool strategies, evaluations. Promotable when skills prove repeatable.
Auto-learning	Every MCP/CLI operation can absorb experience — reasoning traces, skills, growth merge, memory ingest — without the model manually calling write tools.
Codebase mirrors	Mirror-neuron mimicry — observe how repos are written, extract motor programs, match style before patching.
Multiverse cognition	Parallel timeline decisions — spawn branches, stress-test invariants, collapse to governed DNA. Connected MCP models submit reasoning via external_parent_decision.

Everything flows through a single capability funnel (see docs/CAPABILITY_FUNNEL.md):

Observe / Decide
       ↓
Experience Graph
       ↓
Growth Merge
       ↓
Memory Bank
       ↓
Skills (learned + fused)
       ↓
Genomes / DNA

The sacred 6-step loop wraps execution: Experience → Overseer → Parent records reasoning → Steering → Execution → write experience back. The Overseer serves the Parent. The Parent is accountable. The graph is the witness.

---

Using AgentDrive as your framework

When AgentDrive is the substrate your agent runs on — not just a tool it calls occasionally — start every serious task with the framework skill playbook:

framework_session_start(task="...", project_id="my-project")
  → anchor + growth merge + matched learned/fused skills

framework_skill_route(task="...", project_id="my-project")
  → ranked playbooks with when_to_call + invoke_hint

framework_skill_run(name="learned-myproject-mimic-growth-merge-briefing")
  → execute bound op or return SKILL.md body

Readable skill names tell models what was learned:

• learned-{project}-{verb}-{focus} — e.g. learned-openmangos-mimic-growth-merge-briefing
• fused-{project}-{axes} — e.g. fused-openmangos-experience-patterns-skills

Every run_operation may emit auto_learning on the result — check it for new skills, growth merge, and memory ingest.

---

What AgentDrive is not

• A hosted vector DB or chat-memory SaaS
• A drop-in replacement for your editor's built-in context window
• Ready without ~10 minutes of setup (that's what the golden path is for)
• Only AD-Grid / Mission Control — those are advanced layers on top of the Drive

The compounding loop for operators: think → learnings log → drive query → next session's context pack. For models: framework_session_start → work → record_reasoning → auto-learning grows the bench.

---

Start here (~10 minutes)

Install:

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

Golden path:

agentdrive golden-path steps          # numbered commands
agentdrive doctor
agentdrive mcp install && agentdrive mcp doctor
agentdrive golden-path run            # seed → think → learnings → drive query

Step	What it proves
doctor	Local home, config, registry healthy
mcp install	Your AI CLI can call Experience Graph + DNA + memory + skills tools
think	Cited synthesis with gap analysis (not generic chat)
learnings log	Operational memory persists across sessions
drive query	Semantic search over your DNA pool

Full guide: docs/GOLDEN_PATH.md

---

Connect any AI model (MCP)

AgentDrive speaks Model Context Protocol — the same surface Grok, Claude, Cursor, and local models use internally.

agentdrive mcp install
agentdrive mcp doctor
agentdrive mcp config          # or --client claude / cursor / generic

Mandatory first action for any connected model: agentdrive_mcp_catalog() — live tool list with when_to_use, examples, read-only hints, and clone/dev setup guidance.

Key tool families:

Intent	Start here
Starting work (AD is your framework)	framework_session_start → framework_skill_route
What do we already know?	experience_graph_get_context_pack → memory_bank_deep_briefing
Which path should we take?	external_parent_decision (MCP model) or multiverse_parent_decision (local LLM)
How is this repo written?	codebase_register_project → codebase_observe_file → codebase_mimic
Remember this outcome	Harness record_outcome → auto-ingest → graph trace

Docs: docs/MCP.md · docs/FOR_AI_MODELS.md · docs/MEMORY_BANK.md

---

Where your data lives

Local-first. User-sovereign. Everything under ~/.agentdrive/:

~/.agentdrive/
├── config.yaml
├── genomes/                         # Global genome registry
├── skills/                          # User + inherited learned skills
├── codebase-patterns/<project>/     # Mirror-neuron observations
├── learnings/                       # Operational JSONL logs
└── swarms/<swarm_id>/
    └── drive/
        ├── memory_bank/             # memories.jsonl + relations.sqlite3
        └── meta_evolution/          # Experience Graph + multiverse sessions

Swarm-scoped storage means each project or mission gets isolated memory that still compounds within its swarm.

---

Architecture at a glance

┌─────────────────────────────────────────────────────────────┐
│  MCP / CLI / TUI / Harness  (any model, any harness)        │
├─────────────────────────────────────────────────────────────┤
│  Framework playbook  →  Auto-learning  →  Operations      │
├──────────┬──────────┬──────────┬──────────┬─────────────────┤
│ Experience│ Growth   │ Memory   │ Skills   │ DNA / Genomes   │
│ Graph     │ Merge    │ Bank     │ learned  │ (Drive engine)  │
│           │          │          │ + fused  │                 │
├──────────┴──────────┴──────────┴──────────┴─────────────────┤
│  Codebase mirrors  ·  Multiverse cognition  ·  Capabilities │
└─────────────────────────────────────────────────────────────┘

Subsystem map: docs/ARCHITECTURE.md
Capability funnel: docs/CAPABILITY_FUNNEL.md
About & docs site: docs/about.md · docs/index.md

---

Advanced — AD-Grid & Mission Control

Complete the golden path first. AD-Grid is the long-lived intelligence world on top of the Drive — persistent inhabitants, Council governance, real-time Tower observability.

agentdrive grid run --swarm-id stabilization-wave-20260531 --with-tower
# → http://127.0.0.1:8421

Connected models join as first-class inhabitants via agentdrive_register_program — governed programs with full experience_graph_* surfaces and code-agency tools.

Guides: docs/AD_GRID_JOIN.md · docs/AD_GRID_VISION.md

---

Documentation

Doc	Purpose
INSTRUCTION.md	Start here (LLMs) — what, why, how, explain to user
FOR_AI_MODELS.md	Deep reference rules for any connected model
CAPABILITY_FUNNEL.md	How intelligence compounds (single mental model)
MEMORY_BANK.md	Deep memory layer — vault/topic, search, anchor
MULTIVERSE_COGNITION.md	Parallel timeline decisions
SKILLS-LIBRARY.md	Skills bench, inheritance, fusion
GOLDEN_PATH.md	Install → verify in ~10 minutes
MCP.md	Connect Grok, Claude, Cursor, local models

Changelog: CHANGELOG.md

---

License

MIT

---

Intelligence that remembers the shape of what it has become.