Forge binds Oxid-DB, Sensory and Surface into one agent: it remembers, perceives, speaks, and reasons over verified knowledge instead of guessing.
$ oxid forge run "summarise our crisp pale lagers" --groundedForge binds memory, perception, and voice into a single agent: one that doesn't just retrieve and respond, but reasons over what it knows.
Forge doesn't autocomplete an answer. It runs a genuine reasoning loop: every step grounded in Oxid's neurosymbolic graph of logic and vectors, so its conclusions are checkable, not invented.
A plain LLM autocompletes from its training.
Forge reasons over the live knowledge graph, and shows its work.
Oxid-DB remembers, Sensory perceives, Surface speaks. Forge is the agent that binds all three: routing every request to the right faculty and weaving the answers back into one coherent loop.
Recalls grounded facts and reasons over the graph: every answer it returns carries its provenance.
Pulls in new context from text, code, and docs on the fly, perceiving the world as it changes.
Speaks OxQL and natural language, calls tools, and returns answers agents can act on.
Watch a forged agent reason over the knowledge graph: querying, recalling cited facts, and grounding its answer. No black box. Every claim is sourced.
Can a Forge agent write to Oxid-DB inside a read transaction?
To answer this I can't guess: I need two facts from the graph: the isolation level of a read transaction, and whether an agent holds the write capability inside one. Let me query both.
FIND?iso?cap
WHEREReadTxnHAS-ISOLATION?iso
ANDAgentHOLDS?capINReadTxn
write("Brewery", { name: ... })
✗ rejected · capability missing under snapshot isolation
No. Inside a read transaction the agent runs at snapshot isolation, and the write capability is not granted, so the write is rejected before it ever reaches the store.
No retrieval theater. Every step is a real OxQL query, every recalled fact carries its source, and the answer is a function of those cited facts: remove a citation and the answer changes.
Answers cite graph facts: no hallucinated sources.
Think → act → observe, until the goal is met.
OxQL queries and MCP tools, called directly.
Every claim traceable back to its source.
Anthropic, OpenAI, Gemini, or local Ollama.
Runs standalone: no extra services to deploy.
Grounded in your graph, retrieving data you control, traceable to the source, and running on the model you choose.
Research preview · v0.8.5: open and evolving.