feat: optimize memory extraction for concise output and precise retrieval

[lishixiang0705]

Problem

In production, extracted memories average 500-2000 chars per item, causing:

1. Embedding vector dilution — any query fuzzy-matches long content, scores cluster in 0.18-0.21 range
2. Poor retrieval discrimination — relevant and irrelevant items score similarly
3. Context bloat — 5 injected memories can exceed 5000 chars (~3000 tokens) per turn

Solution

1. Prompt optimization (memory_extraction.yaml)

• Add explicit length targets: abstract ~50-80 chars, content 2-4 sentences
• Add good/bad examples showing concise vs verbose patterns
• Guide LLM to split multi-topic memories into separate atomic items
• Emphasize fact-dense "sticky note" style over narrative expansion

2. Vectorization improvement (memory_extractor.py)

• Use abstract instead of content for embedding generation
• Shorter text → more focused vectors → better cosine similarity discrimination
• Fallback: abstract or content ensures no empty embeddings

Expected Impact

• Memory size: 500-2000 chars → 100-300 chars per item
• Injection cost: ~3000 tokens/turn → ~600 tokens/turn (80% reduction)
• Retrieval precision: score spread from 0.03 → 0.15+ between relevant/irrelevant

Files Changed

• openviking/prompts/templates/compression/memory_extraction.yaml — prompt template
• openviking/session/memory_extractor.py — 2 lines: set_vectorize text source

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

lishixiang seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[yangxinxin-7]

Thanks for the work on this — the overall direction makes sense. Two things worth checking before merging:

1. Prompt inconsistency
   The Three-Level Structure section now describes L2 as "2-4 sentences", but the # Few-shot Examples section is unchanged and
   still shows verbose narrative-style content. Since LLMs tend to follow examples more than instructions, the new guidance
   may have limited effect until the few-shot examples are updated to match.

2. Language mixing in examples
   The new ✅ GOOD examples are in Chinese while the rest of the prompt is in English. This may unintentionally bias output
   language for non-Chinese users when output_language is "auto".

[ZaynJarvis]

there are few-shot examples below, these GOOD BAD examples seems redundant or misplaced.

[ZaynJarvis]

these few lines prompt optimization is great. will possibly cherry pick then merge if no further update.

[ZaynJarvis]

Cherry-picked into #1522 for batch merge along with #533 and #951.

What I kept:

• The length targets on the Three-Level Structure section (~50-80 / 3-5 bullets / 2-4 sentences) — this is the core of the prompt optimization.
• The set_vectorize(abstract or content) switch in memory_extractor.py.

What I dropped, addressing review feedback:

• The BAD/GOOD content example blocks — redundant with the Few-shot Examples section below (my earlier inline comment).
• The Chinese-mixed example values — keeps the prompt English-only so output_language: auto doesn't get biased for non-Chinese users (@yangxinxin-7's point).

I did not update the existing Few-shot Examples to match the new length targets (@yangxinxin-7's first point) — that felt like a separate scope change worth handling in a follow-up so this PR stays small.

Memory-quality effect (embedding score spread, retrieval precision) is pending eval on the bundled PR. Thanks @lishixiang0705.

[ZaynJarvis]

Following up on the earlier cherry-pick attempt — closed #1522 and re-did the cherry-pick properly so your original commit is preserved via git cherry-pick (lishixiang authorship intact). Bundled with #533 in #1530 — memory effect eval pending before merge. I also dropped the inline BAD/GOOD content blocks and reverted the L0 bullet examples to the original English strings per earlier review feedback.