v0.1.7 — Smarter Memory

What's new in v0.1.7

🧠 Vault deduplication

Serenity was writing the same memory as 2–3 separate vault files — once during the conversation, again during reflection, sometimes again in a follow-up session. vault_write now checks existing notes for ≥60% title word overlap before creating anything new. If a similar note already exists, new content is appended under a dated section instead of spawning a duplicate file. Your vault stays clean and searchable.

🔁 NNN duplicate entries eliminated

The vector memory store (NNN) was receiving the same facts 2–3 times per session:

• Once from _auto_nnn_encode_direct (runs immediately after any vault_write)
• Again from SESSION_REFLECTION Step 3 (re-encoding the same conversation content)

SESSION_REFLECTION now only stores insights that are novel to the reflection pass — patterns and causal lessons surfaced by looking back at the session, not re-encoding what was already written to vault mid-conversation.

⚡ Context window auto-tunes per model

The hardcoded contextWindowTokens override was bypassing Serenity's built-in model-aware auto-detection. Now removed — context size is set automatically at startup:

Model size	Context window
≤ 4b	20,480 tokens
≤ 9b	28,672 tokens
≤ 20b	40,960 tokens
Cloud	80,000 tokens

No more OOM crashes from over-allocation on small models, no more cramped responses from under-allocation.

---

Full changelog: CHANGELOG.md