Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
89 lines (66 loc) · 3.54 KB

ARCHITECTURE.md

File metadata and controls

89 lines (66 loc) · 3.54 KB

AgentDrive Architecture (Overview)

AgentDrive is a local-first platform that gives AI agents structural, compounding memory — not just retrieval.

For the capability funnel, see docs/CAPABILITY_FUNNEL.md.

For model onboarding rules, see docs/FOR_AI_MODELS.md.

Core subsystems

Layer Location Purpose
Drive engine src/agentdrive/drive/ Swarm-scoped persistence, ingest, query
Operations registry src/agentdrive/operations/registry.py Canonical MCP/CLI surface + auto-learning
Experience Graph src/agentdrive/evolution/experience_graph.py TypedEdges, cycles, fabric reasoning
Growth merge src/agentdrive/learning/growth_merge.py Cross-surface pattern compounding
Memory Bank src/agentdrive/memory/ Deep AI databank (memories.jsonl, relations)
Framework skills src/agentdrive/learning/framework_skills.py Route learned/fused playbooks per task
Multiverse cognition src/agentdrive/cognition/ Parallel timelines → governed collapse
Skills + learning src/agentdrive/skills/, learning/auto_absorb.py Distillation, inheritance, fusion
Codebase mirrors src/agentdrive/codebase/ Pattern recognition + mimicry
Golden path golden_path.py Install → doctor → MCP → verify
MCP adapter adapters/mcp_server.py Any-model tool surface

Naming: Product surface is Drive (AgentDrive, get_swarm_drive_path). The YAML config key pool: and some method names (get_pool_stats) are historical — they refer to the same engine under src/agentdrive/drive/.

Capability funnel

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

Data layout

~/.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/                       # Shared per-swarm Drive (v2)
        ├── ingest.jsonl
        ├── genomes/
        ├── memory_bank/             # memories.jsonl + relations.sqlite3
        └── meta_evolution/          # Experience Graph + multiverse sessions

Swarm model (v2 / Milestone 2a): All sub-agents in a swarm share one Drive at swarms/<swarm_id>/drive/. Sub-agents namespace writes via Genome author tagging (sub:<id>), not separate directories. MCP create_scoped_pool() and SwarmDriveManager use the same path.

Legacy v1 trees (swarms/<id>/<subagent>/pool/) may exist on disk from older sessions; new provisioning always uses the shared drive/ layout.

Integration surfaces

  • MCP — primary for frontier/local models (agentdrive-mcp)
  • CLI / REPL — agentdrive + cli_surface.py shared handlers
  • TUI — default operator shell when no --cli
  • Harness — in-process outcome capture for adapters

See docs/MCP.md, docs/INTEGRATION.md, docs/GOLDEN_PATH.md, docs/MEMORY_BANK.md.

Archived / legacy (do not extend)

  • archive/ui-webpage-legacy-* — superseded web templates
  • archive/development-history/ — stabilization sprint reports
  • src/agentdrive/backup/ — still referenced by tests; migrate before removal

See CHANGELOG.md for recent work.