Knots
Cross-episode synthesis you can recall in one shot.
A knot is a single synthesized narrative over many related episodes. Once an entity in your graph accumulates enough connected memories, Breeth condenses them into one block of prose your agent can pull in a single retrieval — instead of stitching ten episodes together at query time.
Why they exist
A vector store gives your agent ten chunks and asks it to assemble meaning. That's expensive (tokens) and lossy (chunk boundaries cut sentences mid-thought). A knot is pre-computed synthesis over the hub entity, written by Breeth in the background. Your agent retrieves the knot once and gets the full picture.
How they're triggered
Breeth runs the knot synthesizer in the post-write pipeline. It picks the highest-value candidates per period — entities that have crossed an importance threshold (degree, recency, intent density). You don't trigger knots yourself. Breeth decides.
Per-tier cap
Each tier has a monthly knot cap (5 / Hobby, 15 / Starter, 40 / Growth, 100 / Pro, unlimited / Enterprise). Once you hit the cap, knot synthesis silently pauses until next billing period — your writes keep working, you just don't get new knots until the counter resets.
No 429 on knots
Unlike writes / intents / retrievals, hitting the knot cap doesn't fail any request. The cap exists to bound our inference cost; from your perspective the worst case is "no new knots this week" — old knots stay queryable.
How to read them
When you POST /v1/search, results include both regular edge hits and any knot summaries available on the matched entities. The knot is delivered inline as a summary field on the hub node, so your agent gets it as part of the same payload — no extra round-trip.
When they help
- Long-running threads: a candidate that's been discussed for three weeks accumulates dozens of episodes. The knot is one paragraph that captures the through-line.
- Decision archeology: "why did we hire Nandini six months ago?" — the knot on Nandini's node has the answer pre-assembled.
- Onboarding new agents: a new agent session in your team reads the knots to bootstrap context fast.
When they don't
- Low-signal entities: if an entity only appears in two episodes, Breeth won't waste compute synthesizing a knot over it. Just retrieve the two episodes directly.
- Operational data: knots aren't right for high-frequency log-style writes. They're for content that builds a narrative over time.