Sleep, knowledge and the morning briefing

What happens overnight: how experiences become trusted knowledge, how contradictions and forgetting work, and what the morning briefing contains.

Last updated Sep 12, 2026

Tulimoa Memory does its real work while your agent is idle. Once a night, starting at 02:17 UTC, a librarian model goes through everything the day left behind. Your agent never has to tidy up its own memory.

When a memory sleeps#

• Experiences that are at least 30 minutes old and not consolidated yet are up for the night.
• A memory sleeps once at least 5 new experiences are waiting, or when the oldest one has waited more than 3 days. A small memory is not woken up for every single note.
• Each night has a token budget for the librarian that grows with the size of the memory, see the storage fee. Whatever does not fit waits for the next night.
• While a storage fee is unpaid, sleep pauses. Recall keeps working.

What the librarian does#

Related experiences are grouped by meaning. For each group the librarian sees the experiences, with personal data already replaced by placeholders, plus the five closest pieces of existing knowledge, and proposes changes. Every change is checked for format, secrets and scope before it is applied.

• Merge. Duplicates and near-duplicates become one strong piece of knowledge instead of five weak ones.
• Abstract. What happened becomes what it means: a fact, an entity, a preference, a procedure, a concept or an open problem.
• Link. New knowledge is connected to what it relates to, is part of, is about, caused or came after, so tomorrow one cue pulls the whole thread.
• Replace. Outdated knowledge is superseded instead of piling up next to the new version.

From provisional to trusted#

• New knowledge starts provisional. It becomes trusted once it was seen at least 3 times and is at least 7 days old.
• Contradictions go into quarantine. When provisional knowledge contradicts trusted knowledge, it is flagged, left out of recall and know, and listed under Needs review in the briefing. A single poisoned note cannot overwrite what the memory has confirmed.
• Forgetting is a feature. Knowledge loses weight over time, with a half-life of about 30 days, unless it is reinforced; every recall hit strengthens it. Provisional knowledge seen fewer than 2 times is removed after 21 days.
• Raw experiences move on. Consolidated experiences go to the archive after 14 days. Non-durable ones are deleted from the live memory after 30 days; what remember and /import store is durable and stays.

What else grows overnight#

• Recipes. Sequences of 2 to 4 successful tool calls that repeat across at least 2 sessions become a procedure the agent can reuse. Recipes come from tool results sent through /ingest with payload.tool.
• Self-model. Once there are at least 5 pieces of knowledge, the librarian writes down what the agent mainly works on, how the user likes things done and where it runs into trouble. Read it with tulimoa__self.
• Gaps. Topics that are named but thinly connected show up in tulimoa__gaps and in the briefing as worth exploring.
• Ideas for open problems. For up to 3 open problems the librarian suggests a next step. These ideas go into the briefing only, never into the knowledge graph.
• PII watch. If a name slipped in without a placeholder, the librarian reports it and the memory replaces it with a placeholder in experiences, knowledge and open loops, up to 10 names per night. The briefing tells you when that happened.

The morning briefing#

After every night with work there is a briefing, dated by the UTC day. It is built from what the night actually did, not written freely by a model, so it cannot make things up. The headline reads like While you were away: 4 new, 2 reinforced, 1 forgotten from 23 episodes.

• For the agent: new knowledge, open problems, warnings, what needs review, ideas for open problems and the PII watch note. Read it with tulimoa__briefing; the headline is also part of the waking state on connect.
• For people: the same plus a list of topics worth exploring, on the /briefing page and in the dashboard.

Archive and timeline#

The knowledge graph keeps what things mean. When you need what literally happened, tulimoa__timeline reads the raw history, and for ranges older than 14 days it reads the cold archive transparently, up to 62 days per request.