Review: AAAK Expand for Pre-Embedding Semantic Quality

Good to see this from you again, @Nitrogonza9 — we reviewed your #433 (contradiction detection) and #434 (auto-KG) previously. This one addresses a real measured problem (84.2% vs 96.6% R@5 for AAAK vs raw on LongMemEval).

expand() Method

The reconstruction logic is solid. Walking the decoded structure and reassembling human-readable fragments is the right approach for embedding quality. Specific notes:

Entity code reversal:

for name, code in self.entity_codes.items():
    if not name.islower() and code not in code_to_name:
        code_to_name[code] = name

The not name.islower() filter ensures you prefer proper names ("Alice") over lowercase aliases. Good heuristic. The code not in code_to_name gives first-registered-name priority, which is reasonable.

Emotion expansion — Using _REVERSE_EMOTIONS built at module level is efficient. The handling of combined emotions ("determ+hope") by splitting on + is correct.

⚠️ Edge case — field parsing ambiguity:
The expand() method iterates zettel fields and tries to classify each one (entity, quote, emotion, topic). But what if a topic contains +? The check:

elif "+" in field and all(f.strip() in _REVERSE_EMOTIONS or f.strip().isupper() for f in field.split("+"))

would match a topic like "TCP+UDP" (both are uppercase) and treat it as combined emotions, yielding nonsense expansions. Consider checking against _REVERSE_EMOTIONS first and only treating it as combined emotions if at least one part is actually an emotion code.

looks_like_aaak() Heuristic

Clean implementation. The "pipe-separated with digit-colon prefix on first field" heuristic should have very low false-positive rates on normal text. The early "|" not in text bailout is good.

MCP Integration

⚠️ Import inside function body:

def tool_diary_write(...):
    from mempalace.dialect import Dialect
    _dialect = Dialect()

This creates a new Dialect instance on every diary write. If diary writes are frequent (our agents write ~50 per cycle), this is wasteful. Consider moving the import and instantiation to module level, or at minimum cache the dialect instance.

Metadata preservation — Storing the original AAAK in aaak_compressed metadata while embedding the expanded text is the right design. It preserves lossless access to the compressed form while giving the embedding model better input.

Impact Assessment

If this improves R@5 from 84.2% closer to 96.6% for AAAK entries, it's a meaningful quality improvement for anyone using the compress workflow. The change is backward-compatible (plain text entries pass through unchanged), and the looks_like_aaak() gate prevents false expansions.

Tests are comprehensive — 12 new tests covering expand, roundtrip, and heuristic detection. The test_expand_roundtrip_from_compress test is particularly valuable.

Good contribution. The TCP+UDP edge case and the per-call Dialect instantiation are worth addressing.

🔭 Reviewed as part of the MemPalace-AGI integration project — autonomous research with perfect memory. Community interaction updates are posted regularly on the dashboard.