Research-only analysis of MemPalace (MemPalace/mempalace, 52.7k stars,
verbatim-storage AI memory over ChromaDB) in the same format as
issues-basic-memory.md / issues-agentmemory.md.
Key conclusions captured for later revisit:
- MemPalace's dominant open-issue cluster (concurrent-write HNSW/FTS5
corruption, filed-but-not-persisted, repair-destroys-data, embedding
drift) maps ~1:1 onto ai-memory invariants #2/#3/#8/#10/#17.
- Worth a brainstorm: a verbatim semantic-recall fallback tier beneath
the compiled wiki (lossless recall without ChromaDB's corruption
surface); honest LongMemEval (M10); deterministic resume-safe ingest;
temporal KG triples.
- Gotchas to avoid in v0.2 vector work + a supply-chain/impostor-domain
note for when the project gets popular.
No ai-memory code or design changed — pending brainstorm.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
AkitaOnRails