chore: remove all external project references from docs and code comments

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

Author: claudioemmanuel

README.md

@@ -366,56 +366,6 @@ tail_preserved=20
 
 ---
 
-## squeez vs MemPalace
-
-MemPalace ([milla-jovovich/mempalace](https://github.com/milla-jovovich/mempalace)) is the closest comparable project — Python, ChromaDB, semantic memory, MCP server, self-described "highest-scoring AI memory system ever benchmarked." We audited its code and benchmarks. Here is what we found.
-
-### Architecture comparison
-
-| Dimension | squeez | MemPalace |
-|-----------|--------|-----------|
-| **Language / runtime** | Rust, single static binary (703 KB) | Python 3.10+, ~80 MB ML stack |
-| **Dependencies** | `libc` only (Unix signal fwd) | `chromadb`, `sentence-transformers`, `pyyaml`, `sqlite3`, … |
-| **Memory model** | JSONL session log + finalized summaries | ChromaDB (vector) + SQLite KG (graph) |
-| **Redundancy** | Exact FNV-1a + fuzzy MinHash trigram Jaccard | None — every write creates a new embedding |
-| **Compression** | 4-stage pipeline (filter→dedup→group→truncate) | AAAK dialect (ad-hoc abbreviations, **lossy**) |
-| **Semantic recall** | Cross-session JSONL summaries (text) | ChromaDB cosine similarity (embeddings) |
-| **MCP tools** | 6 read-only, hand-rolled JSON-RPC 2.0 | 19 tools via fastmcp |
-| **Test coverage** | 35 integration test files, 287 tests | 4 test files, 0 AAAK/compression tests |
-| **Latency** | < 0.3 ms p50 filter mode | Not reported (embedding inference ~50–200 ms) |
-| **Token reduction** | Up to 95%, 92.8% aggregate | AAAK: ~30–40% (ad-hoc, no systematic measurement) |
-
-### Benchmark audit
-
-MemPalace claims 96.6% R@5 on LongMemEval-S. We replicated the benchmark methodology from `benchmarks/`:
-
-| Claim | Reality |
-|-------|---------|
-| **96.6% R@5** | Achieved with **raw ChromaDB** — no palace structure, no AAAK encoding |
-| **AAAK encoding** | Degrades to **84.2% R@5** (−12.4 pp vs. raw) |
-| **Rooms mode** | Degrades to **89.4% R@5** (−7.2 pp vs. raw) |
-| **+34% from palace structure** | Numbers (60.9 / 73.1 / 94.8 with 22,000 memories) **do not appear in benchmarks/**; no reproducer |
-| **100% R@5** (claimed) | Teaching-to-test: 3 hand-crafted questions whose answers are in training. Honest held-out = 98.4%. |
-| **LoCoMo 100%** | Gamed: top-k=50 exceeds 19–32 sessions per conversation, so ground truth is always in the candidate pool. |
-
-**Summary:** the ChromaDB embedding layer genuinely scores well on long-context retrieval benchmarks. The palace metaphor (wings/rooms/halls/closets/drawers/tunnels) is ChromaDB metadata — it degrades accuracy. AAAK compression degrades accuracy further.
-
-### What we adopted from the audit
-
-The gap analysis informed squeez 0.3+:
-
-| MemPalace technique | Adopted in squeez | Notes |
-|---------------------|-------------------|-------|
-| Fuzzy semantic similarity | ✅ MinHash trigram Jaccard (no embeddings) | Zero-dep compliant; ≈ same dedup power for repeated CLI output |
-| Temporal validity on memories | ✅ `valid_from` / `valid_to` on `Summary` | Summaries age from `valid_to` when invalidated |
-| Benign-output threshold scaling | ✅ 2× threshold when no error markers | Successful builds stay verbatim |
-| Auto-teach protocol payload | ✅ `squeez protocol` / `squeez_protocol` MCP tool | LLM learns markers on first call |
-| MCP server for session query | ✅ 6 tools, hand-rolled JSON-RPC 2.0 | No fastmcp/mcp.server dep; `libc`-only intact |
-| Embeddings / vector store | ❌ — would require ~80 MB ML stack | Out of scope for zero-dep constraint |
-| Graph KG (SQLite) | ❌ — multi-hop QA not needed for CLI session context | Scope mismatch; summaries cover the use case |
-
----
-
 ## Platform notes
 
 ### OpenCode