Option
memory of agent
The table that holds the conversations of the agent.
Value
{ in, thread, of, keep, after, shrink, model, recall }, of which in is required. Left out, a run uses no memory.
shrink: { nearest: <count> } sends of the older turns of the conversation, beside the latest keep, the <count> nearest by meaning to the newest message, each turn whole and in its place. recall: <count> adds, before each model step, one message of the engine's own: so many older messages of the starter, of every conversation the starter may read, nearest by meaning to the newest one, which are not sent in full by keep and shrink. For both, the memory table states a vector column that the engine fills from text, as meaning: vector({ size: 1536, model: embedder, by: 'text' }); the newest message is turned into a vector by that column's model in the step recall, once a run. The turns of the conversation are its own, as every turn is sent; the rows that recall brings are found with near under the table's read rule and held to the rows the starter owns, so nothing of another person is recalled. A table with no such column is code.T7007.
Facts about the person are told by knows (option.agent.knows), and written by a change of the builder's own given to the agent as a tool.
The engine's own loop reads and writes the memory table: a driver of your own (adapter.Driver) writes the rows of its runs itself, as code.T7007 gives them.
Example
export const helper = agent('helper', {
model: embedder,
instructions: 'You help with notes.',
tools: [countNotes, { tool: nightly, tries: 2, about: 'Tidies up at once', line: false, name: 'tidyNow' }],
limits: { steps: 12, perRun: { tokensOut: 50000 }, time: '10 minutes', ask: '7 days', result: '1 MB', toolCalls: 100, toolsAtOnce: 8, pass: '5 seconds' },
memory: { in: notes, thread: 'team', keep: 10, after: 10, shrink: 'summary', model: embedder },
})Errors
code.T7007: the memory table of recall or of shrink: { nearest } holds no vector column filled from text.