agent settings

An agent's settings are one object, the same whether you set them in the dashboard or with PATCH /agents/{agent} ({ "settings": { … } }, merged into what is there). A list (memory.kinds, memory.exclude, person.never) replaces the list it is given for, it is never added to it. Every setting has a default, so an agent works the moment it has a name. A key or a connected app may lower what an agent reads of you; only you, in your agents space, can widen it.

memory: what it keeps#

setting	default	what it changes
`memory.infer`	`true`	pull separate memories out of what it is told, or keep the text as given
`memory.instructions`	empty	what to keep, in your words (up to 4,000 characters)
`memory.kinds`	all six	the kinds it keeps; anything else is skipped
`memory.exclude`	none	topics it never keeps (up to 20)
`memory.redact`	secrets and cards on	emails, phone numbers, keys and passwords, card numbers: hidden before anything is stored or sent to a model
`memory.retention`	episodes 90 days, runs 30 days	days until a memory is deleted: `episode_days`, `run_days`, `default_days` (`null` is never)
`memory.max_memories`	10,000	counts earlier versions too: earlier versions go first, then the oldest, least important memories. An earlier version is deleted 180 days after it was replaced
`memory.dedupe`	`merge`	a near-duplicate becomes a new version of the old one, or (`keep`) both stay

person: what it learns about you, and reads of you#

setting	default	what it changes
`person.funnel`	`ask`	`ask` waits for your yes, `auto` files at once, `off` keeps it to the agent
`person.min_confidence`	0.7	how sure it must be that a memory is about you before it is offered
`person.never`	your body and health, your circles, the people in your life, what stays with you	the parts of your life it may never file under
`person.disclosure`	`private`	the most open anything it files may be; your memory can only make it more private
`person.read`	`standard`	what it may read of you: `none`, `standard` (what any connected app can), `personal`, `private`

recall: how it remembers#

setting	default	what it changes
`recall.limit`	10	how many memories come back (1–100)
`recall.threshold`	0	the lowest score that still comes back (0–1)
`recall.rerank`	`false`	re-read the best matches by meaning, a little slower
`recall.recency`	0.3	how much newer beats older (0–1)
`recall.include_person`	`false`	bring what it may read of you into every recall

limits#

setting	default
`limits.writes_per_minute`	120
`limits.reads_per_minute`	600
`limits.extract_per_day`	1,000

limits.extract_per_day: 0 turns reading off: a call with infer on answers 400, and infer: false still keeps the text as given. All of one person's agents share 20,000 extractions a day. A 429 means a limit was reached; when the limit cannot be checked just then, the text is kept as given and skipped says so. Listing and reading agents are counted against the read limit too.

an example#

bash

curl -X PATCH https://innernet.live/api/v1/agents/inbox-keeper \
  -H "Authorization: Bearer $INNERNET_KEY" \
  -d '{ "settings": {
        "memory": { "instructions": "keep who writes, what they want, and by when", "retention": { "episode_days": 30 } },
        "person": { "funnel": "ask", "never": ["relationships", "hauntings", "kin", "somatic", "aversions"] } } }'