neruva-goose 0.1.0

Goose extension for Neruva agent memory + reasoning substrate. Wires Neruva's three auto-pilot layers (route intent / reflect / extract) into Goose's pre-message and post-tool events. Pattern-C: substrate emits the prompt, your LLM runs it, results pushed back. Substrate stays $0/call. Pairs with the @neruva/mcp Goose MCP extension for full tool surface.

neruva-goose

Goose (Block Inc) extension for the Neruva agent memory + reasoning substrate.

Two integration paths -- use one or both:

Path 1: @neruva/mcp as a Goose MCP extension (zero Python)

Drop the Neruva MCP server straight into Goose's extension config. You get every Neruva substrate tool (records, KG, causal, analogy, CBR, ToM, EFE planning, code graph, replay) inside Goose with no glue code.

~/.config/goose/config.yaml:

extensions:
  neruva:
    type: stdio
    cmd: npx
    args: ["-y", "@neruva/mcp@latest"]
    env:
      NERUVA_API_KEY: nv_your_key_here

Restart Goose. Run goose mcp list -- you should see ~80 Neruva tools.

Path 2: neruva-goose Python package (auto-pilot)

For pattern-C auto-pilot (route intent / reflect / extract running through your LLM, with substrate emitting only the prompts), install this Python package and wire its hooks into your Goose extension.

pip install neruva-goose

from neruva_goose import NeruvaAutoPilot, NeruvaAutoPilotHooks

ap = NeruvaAutoPilot(
    api_key="nv_...",
    namespace="my_session",
    llm_callable=lambda prompt: my_llm.complete(prompt),
)
hooks = NeruvaAutoPilotHooks(ap, reflect_every=10)

# Pre-message: classify intent so you can pick the right Neruva tool
intent = hooks.before_user_message(user_text)
print(intent)
# -> {"primary_intent": "counterfactual", "confidence": 0.87,
#     "suggested_tool": "agent_counterfactual_rollout", ...}

# Post-tool: auto-extract facts to the auto-KG, fire reflection every 10
hooks.after_tool_call("agent_recall", str(tool_output))

# Optional: buffer assistant turns
hooks.after_assistant_message(assistant_text)

# End of session: force a final reflection
hooks.after_session(force=True)

The three layers:

Layer	Hook	Substrate prompt	What happens
1	after_tool_call	inline extraction prompt	Triples extracted from tool output, pushed to <ns>-auto-kg.
2	before_user_message	/v1/agent/route_intent_prompt	Returns top intent + suggested Neruva tool.
3	after_session / every N tool calls	/v1/agent/reflect_prompt	Decisions / facts / mistakes / open_questions written to substrate.

Pattern-C, $0/call

The substrate never runs your LLM. It hands back canonical prompts; your code runs them; structured results come back via the existing client. That keeps the substrate deterministic and free at the call boundary, and means your LLM provider, model choice, and rate limits are entirely yours.

License

MIT. Same as neruva-mcp, neruva-record, neruva-langchain, neruva-langgraph, neruva-crewai.