Pick up the work.
Store checkpoints, current status and run identifiers so the runtime can load the state it needs for its next step.
AI agents
Give the next step
something to remember.
Keep run state, useful memories and tool outcomes in one database. Retrieve the context your agent needs and inspect the records behind its work.
Context across steps
Keep task history and source records within reach of your agent.
Illustrative applicationWhat you can build
Make the next turn better informed.
Recall useful context.
Retrieve related memories by meaning, or find a specific observation by its words and structured fields.
Inspect what happened.
Connect observations and tool events to their originating run and source records using stable identifiers.
The application workflow
Recall. Act. Record.
Your runtime controls the agent loop. The database holds the state and evidence the runtime reads and writes along the way.
Suggested record structure
An application-owned memory
-
memory_id -
agent_id -
session_id -
observation -
source_id
Your schema, built around your application.
-
Load the current context
Read run state and retrieve relevant memories. Resolve the returned IDs to source records before preparing the next prompt.
-
Let the runtime choose the action
The application calls models and tools, checks permissions and obtains approval where the workflow requires it.
-
Save the outcome
Write tool events, new observations and the updated run state. Track separate vector or text-index writes before treating a memory as ready for retrieval.
How it fits together
The data in one place.
The workflow in your hands.
Explore the architecture 
OriginChainDBStore and retrieve.
Stores runs, memories, events and declared relationships. Provides structured lookup, similarity and keyword retrieval for the context your runtime needs.
Coordinate and decide.
Chooses tools, calls models, manages retries and decides when to act. Carries agent and session scope into every request.
ASK can serve as a scoped query tool that returns database rows and an executed plan. It does not orchestrate an agent or approve external actions.
Build for real use
Start with a workflow.
Check the whole result.
How will repeated tool calls be handled?
Use stable run, memory and tool-call IDs. Make retries explicit so retrying storage work does not repeat an external action.
What happens when a memory expires?
Coordinate removal of the source row and application-managed vector or text entries. Define which representations exist and track each removal.
Can one session retrieve another's context?
Apply database access controls and validate agent/session scope in the service. Retrieval metadata filters help narrow results but do not replace authorization.
Build with OriginChainDB