feat(prompts): add facet guard and length limits to memory_merge_bundle

[lishixiang0705]

Problem

The current memory_merge_bundle.yaml template has three issues that degrade memory quality over time:

1. No facet coherence check — Two memories sharing the same category (e.g. both preferences) are always merged, even when they describe completely unrelated topics (e.g. "Python code style" + "food preferences"). This produces bloated, semantically diluted memories.

2. No length constraints — Merged content grows unbounded. In production, single memory items can exceed 1000+ characters, causing:

   • Embedding dilution (vector search scores cluster in a narrow band with ~0.02 spread)
   • High token cost when memories are injected into LLM context
   • Low retrieval precision

3. Accumulate-only strategy — The template instructs to "keep non-conflicting details", which means memories only grow, never condense. Old, superseded facts persist alongside new ones.

Solution

Upgrade template to v2.0.0 with three changes:

1. Facet guard (decision: skip)

Before merging, the LLM must verify both memories describe the same specific facet/topic. If they cover different facets, output {"decision": "skip"} to keep them separate. Includes concrete examples for both cases.

2. Hard character limits

• abstract: ≤ 80 characters
• overview: ≤ 200 characters
• content: ≤ 300 characters

When limits would be exceeded, aggressively summarize — drop older details first, preserve specific values (names, numbers, versions) over narrative.

3. Condensed snapshot strategy

Replace "keep non-conflicting details" with:

• Conflicts → keep newer version only
• Non-conflicts → condense to essential facts
• Result should read as a clean, up-to-date snapshot — not a changelog

Breaking Changes

• The decision field now accepts "skip" in addition to "merge". Callers that parse the merge output must handle this new value.
• Merged content will be significantly shorter than before. Downstream systems that rely on verbose L2 content may need adjustment.

Testing

Tested in production with ~2800 vectors and ~76 leaf memory files. The facet guard correctly prevents cross-topic merges, and length limits keep individual memories within embedding-friendly bounds.

[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.}

[ZaynJarvis]

Cherry-picked into #1522 for batch merge along with #549 and #951. The facet guard + length limits + condensed-snapshot strategy are all preserved; I tightened the prompt body a bit (39 changed lines vs. the original 64) but the three behavioural changes are identical.

Memory-quality effect is pending eval — I'll run retrieval precision / score-spread numbers on the bundled PR before merging. Thanks for the work @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 as-is via git cherry-pick (author attribution intact). Prompt-bundle changes (#533 + #549) are now in #1530 — memory effect eval pending before merge.