How fast is context assembly?

Under 100ms in production. Three layers fetched in parallel with 3-second per-layer timeouts, scored, and injected into the system prompt before the model starts generating.

What is the priority order?

Soul blocks first (identity is non-negotiable), then memories, then episodes, then procedures. Soul always fits; the rest compete for the remaining token budget by score.

How is the token budget decided?

Per plan. Starter and Pro: 4K tokens. Max: 8K tokens. The budget caps how many memories/episodes/procedures land in the prompt — the highest-scoring ones win.

What if a layer fails to fetch?

Each layer has a 3-second timeout. If a layer times out (e.g. semantic search degraded), the assembler proceeds with what it has — the conversation never blocks waiting on a slow tier.

Context Assembly

Context Assembly is the step that turns your stored memory into a working system prompt for every conversation. It fetches Soul blocks, relevant memories, recent episodes and applicable procedures in parallel, scores each candidate by five factors and assembles the result within the model's token budget — all in under 100ms. Priority order: Soul → memories → episodes → procedures.

Context Assembly is the process that builds the complete system prompt for every conversation. It intelligently combines soul blocks, relevant memories, recent episodes, and applicable procedures within a token budget. Prioritized: soul → memories → episodes → procedures.

See plans