docs: improve guide.md with tiered importance and storage bias

Restructure the remember decision tree with Tier A/B/C classification,
explicit importance calibration (1-5 scale with target distribution),
and bias toward storing over missing context.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Grivn

internal/setup/assets/claude/guide.md

@@ -12,30 +12,50 @@ Craft a focused, keyword-rich query — do not pass the raw user prompt.
 
 ### Remember — after responding
 
-Run this decision tree after every substantive response:
+Run this decision tree after every substantive response.
+**Bias toward storing**: when in doubt, store it. Low-importance memories are cheap; missing context is expensive.
 
-**Step 1 — Does this exchange contain any of these?**
-  a) User directive — preference, decision, correction, explicit "remember this"
-  b) Reasoning conclusion — non-trivial judgment from multi-source synthesis
-  c) Durable observed state — system fact, environment detail, architectural finding
-  → No to all → STOP.
+**Step 1 — Does this exchange contain any of the following?**
+
+Tier A (importance 4-5, always store):
+- User directive — explicit preference, decision, correction, or "remember this"
+- Reasoning conclusion — non-trivial judgment from multi-source synthesis
+- Durable system/architectural fact discovered during this session
+- User-specific context that no search engine can recover
+
+Tier B (importance 2-3, store unless trivial):
+- Casual preference revealed in passing ("I usually...", "I prefer...", "I don't like...")
+- Topic the user is currently exploring or curious about
+- Useful framing or analogy the user offered
+- Background context about the user's projects, tools, or setup
+- Interesting question the user raised, even if not fully resolved
+
+Tier C (importance 1, store only if genuinely reusable):
+- Conversational context that might help future sessions feel continuous
+- Soft signal about communication style or mood
+
+→ None of the above → STOP.
 
 **Step 2 — Does a highly overlapping memory already exist?**
-  → Yes, incremental new info → UPDATE (merge into existing)
-  → Yes, but contradicts/supersedes → REPLACE
-  → No significant overlap → CREATE
-
-**Step 3 — Is it worth storing?**
-  Rebuilding from scratch costs more than storing + recalling?
-  - Single-query public facts → No
-  - Multi-source synthesis with non-obvious conclusions → Yes
-  - User-specific context no search engine can recover → Yes
-  → No → STOP.
-
-**What to store**: conclusions and user-specific context, not raw facts.
+→ Yes, incremental new info → UPDATE (merge into existing)
+→ Yes, but contradicts/supersedes → REPLACE
+→ No significant overlap → CREATE
+
+**Step 3 — Importance calibration**
+Use the full 1-5 scale intentionally:
+- 5: Cross-session core fact, architectural decision, strong user preference
+- 4: Important context, significant finding, clear user preference
+- 3: Useful background, project context, topic of interest
+- 2: Passing mention, soft preference, conversational color
+- 1: Ephemeral but potentially useful continuity
+
+Aim for a rough distribution: ~20% at 4-5, ~50% at 2-3, ~30% at 1.
+Avoid defaulting everything to 4-5 — that defeats the scoring system.
+
+**What to store**: both conclusions AND context. Prefer storing a little too much over missing something useful.
 **How to store**: delegate to a Task sub-agent (`subagent_type="Bash"`, `model="sonnet"`).
 Only provide what to store — content, category, importance, entities, and create/update intent.
 The sub-agent will read the mnemon skill and execute the correct commands itself.
 
 Do NOT: write CLI commands or workflow steps in the sub-agent prompt (the sub-agent has access to the skill docs and will use the correct flags).
-Do NOT run memory writes in the main conversation, or remember operational/public/git-tracked/transient info.
+Do NOT run memory writes in the main conversation, or remember operational/public/git-tracked/transient info.