Limitations¶
The current release is an evidence-first alpha, not a production memory service. Treat these as design constraints, not fine print.
API and storage stability¶
- API and on-disk schemas may change before 1.0.
- Schema 2 has an explicit v0.1 → v0.2 migration plus verified backup/restore, but this is not a general downgrade or arbitrary-version migration framework.
- The storage root must be dedicated to Citefold. Unrelated files in an uninitialized non-empty directory are rejected rather than adopted.
- Candidate list/approve/reject exists in both Python and CLI paths; a bulk or graphical review workflow is not included.
- Pin/unpin is available for active records; pinning only exempts decay and does not guarantee recall or prevent correction, archival, or deletion.
Language and extraction¶
- Local text capture and recall are language-agnostic at the evidence level.
- The built-in deterministic direct-write parser recognizes only a narrow set of explicit Chinese preference/reminder phrases.
- Arbitrary semantic extraction needs supplied candidates or optional model consolidation.
- Multi-Episode consolidation currently batches selected observations; entity/time/topic grouping and novelty/contradiction quality have not been independently evaluated.
Retrieval and budgets¶
- The default path uses lexical matching and SQLite FTS5; optional embeddings are a secondary, rebuildable signal.
token_budgetis a logical character proxy, not an exact provider token count.- LongMemEval retrieval is strong, but end-to-end QA is materially lower and uneven by question type.
- Results have not been independently reproduced by a third party.
Multimodal quality¶
- The offline benchmark uses supplied observations and does not measure OCR, ASR, vision, codec, or reader-LLM quality.
- OpenRouter endpoint availability and ZDR routing can change.
- Audio normalization and video extraction depend on external
ffmpeg/ffprobeexecutables. - Video uses scene frames and a conservative short-clip fallback; it is not general video understanding.
- Missing provider confidence is stored as
0.0, so such output will not silently become stable memory.
Security and privacy¶
- Citefold has no authentication, remote authorization service, or application-level encryption at rest.
- It is not a sandbox for malicious code in the same Python process.
- JSONL ledgers are auditable but not cryptographically tamper-proof.
- ZDR routing is not a compliance certification.
- Hard deletion does not erase backups, exported logs, or external provider copies.
Operations and scale¶
- Root shared/exclusive locking and scope/ledger serialization are designed for local POSIX process/thread coordination; network-filesystem semantics remain unverified.
- Windows multi-process locking parity is not established.
- v0.1 processes do not participate in the schema 2 root lock; they must be stopped before migration and must not be restarted against the migrated root.
- Migration preflight, verified ZIP backup, additive recovery state, legacy concurrent-change preservation, and journaled restore are covered by local deterministic tests, including a store generated by v0.1.0 and key breaks before/after manifest commit and between directory swaps. This is not exhaustive crash injection, a real power-loss test, a production recovery drill, or proof across arbitrary filesystems.
restore --replaceintentionally retains the prior root as a sensitivedisplaced_root; its sibling intent journal also makes the root parent operational state. Operator review, retention, and deletion are not automated.- Distributed replication, sharding, high availability, and multi-node consistency are not implemented.
- There is no published scale, latency, storage-growth, or cost envelope yet.
- SQLite and filesystem projections are appropriate for embedded use, not a claim of high-throughput service readiness.
Biological analogy¶
Citefold implements explicit episodes, consolidation, cues, reinforcement, decay, and forgetting. It does not reproduce emotional modulation, sleep-dependent consolidation, reconstructive recall, implicit memory, sensorimotor learning, or consciousness. “Human-memory-inspired” describes design prompts, not measured equivalence to the brain.
Evaluation gaps before 1.0¶
- real users over weeks or months;
- real OCR/ASR/video quality and error propagation;
- adversarial media and prompt-injection evaluation beyond deterministic fixtures;
- large-scale concurrency, exhaustive crash/power-loss injection beyond the covered deterministic breakpoints, and production recovery drills;
- independent security/threat-model review;
- migration compatibility beyond the checked-in v0.1 → v0.2 fixture and with real deployment histories;
- third-party benchmark reproduction.
If any of these are hard requirements, treat Citefold as a component to evaluate—not a finished solution.