diff --git a/build-an-oracle/reference/bundled-plugins/memory.mdx b/build-an-oracle/reference/bundled-plugins/memory.mdx
index 4427975..73f8954 100644
--- a/build-an-oracle/reference/bundled-plugins/memory.mdx
+++ b/build-an-oracle/reference/bundled-plugins/memory.mdx
@@ -51,14 +51,16 @@ Memory context does not arrive in the prompt because the agent calls a tool to f
 5. `relationships` — people and organizations mentioned in past conversations
 6. `recent` — notable things from the last few sessions
 
-**When it runs:** at agent-compile time, before turn 1. The fetcher runs once per session and caches the result for **5 minutes** (keyed by `sessionId`). Subsequent turns within the same session reuse the cached context — the Memory Engine is not called again unless the cache expires.
+**When it runs:** at agent-compile time, before turn 1. The fetcher caches the result for **5 minutes**, keyed by `roomId`. Subsequent turns within the same room reuse the cached context — including a fresh session for the same room — and the Memory Engine is not called again unless the cache expires.
 
 **What appears in the prompt:** if at least one slot is non-empty, the runtime inserts a `## What you know about the user` block into the system prompt containing all populated slots. If every slot is empty (no prior memory for this user), the block is omitted entirely.
 
+The block is deduped and budgeted before it reaches the model: entities are collapsed by name (keeping the richest summary), facts / episodes / communities are collapsed by normalized text so exact repeats across the six slots don't render 4-5 times, and per-entity and whole-block caps are applied. If either cap trims content, the block ends with a `_(More remembered — ask me to recall.)_` footer — trimmed detail is still reachable via the memory tool. No unique fact is dropped, and paraphrases survive.
+
 **Implication for oracle authors:** you do not need to instruct the agent to "look up the user's context" or "recall memory before responding" in `config.prompt.opening` or anywhere else. The context is already in the system prompt when the agent sees the user's first message. Adding such instructions is redundant and wastes tokens.
 
 <Note>
-The 5-minute session cache means very-recent memory writes (e.g. the agent just called `memory-engine__add_memory` in the same session) may not appear in the fetched context until the next session or cache expiry. This is intentional — the fetcher is optimised for read latency, not write-through consistency.
+The 5-minute room-keyed cache means very-recent memory writes (e.g. the agent just called `memory-engine__add_memory` in the current turn) may not appear in the fetched context until the cache expires. This is intentional — the fetcher is optimised for read latency, not write-through consistency.
 </Note>
 
 ## Adding global oracle knowledge
diff --git a/build-an-oracle/reference/environment-variables.mdx b/build-an-oracle/reference/environment-variables.mdx
index b3ed7cf..f2db1e6 100644
--- a/build-an-oracle/reference/environment-variables.mdx
+++ b/build-an-oracle/reference/environment-variables.mdx
@@ -73,6 +73,19 @@ Always present. Both are `z.coerce.number()` — set them as plain integer strin
 
 Cross-field check (`validateLlmProviderKey`): the API key for the selected `LLM_PROVIDER` must be present — `OPEN_ROUTER_API_KEY` when `LLM_PROVIDER=openrouter` (the default), `NEBIUS_API_KEY` when `LLM_PROVIDER=nebius`. A missing key fails boot with a named-field error (e.g. `OPEN_ROUTER_API_KEY` / `NEBIUS_API_KEY`) rather than a generic upstream 401 at request time. The per-role model ids are hardcoded per provider — there is no env var to swap the main model id (use the `resolveModel` hook for that).
 
+### Latency tuning
+
+| Variable | Type | Default | Notes |
+| --- | --- | --- | --- |
+| `MAIN_REASONING_EFFORT` | `'low' \| 'medium' \| 'high'` | `'medium'` | Extended-thinking effort for the **main** agent role only. Sub-agent reasoning is unchanged. Lower values reduce time-to-first-token at some cost to hard multi-step reasoning; higher values do the reverse. `'medium'` preserves current behaviour, so leaving this unset is safe. |
+
+<Note>
+Two related latency wins ship **always-on** — there is no env var to disable them:
+
+- **Request-path caches.** Three wins ship together. The per-user `SqliteSaver` is reused across the two build-time hook calls per turn instead of being rebuilt on each call. The agent builder reads prior build-time checkpoints via a light path that skips the message-history join (O(1) instead of O(history)). The Memory Engine `userContext` cache is keyed by `roomId` instead of `sessionId`, so a fresh session for the same room reuses the cached context.
+- **Memory-context compaction.** The `## What you know about the user` prompt block is deduped across all six memory buckets (entities collapsed by name, keeping the richest summary; facts / episodes / communities collapsed by normalized text — lowercase, punctuation stripped, whitespace collapsed) and bounded by fixed internal caps (per-entity summary and whole-block budget). When either cap trims content, the block ends with a `_(More remembered — ask me to recall.)_` footer so trimmed detail stays reachable via the memory tool. No unique fact is dropped; paraphrases survive.
+</Note>
+
 ### Misc
 
 | Variable | Type | Default | Notes |