Qorx

Qorx Reference Papers And External Sources

Qorx is not built from vibes. The current repo keeps a local PaperQA corpus in research/papers, plus official provider and provenance references in the docs. This file is the readable map.

The papers support the architecture class around Qorx: retrieval-backed context omission, prompt compression, repository memory, cache-aware request design, agent memory, and runtime cache/KV scope. They do not prove that Qorx wins on every task. Qorx-specific claims still need Qorx-specific benchmarks.

Prompt Compression

Reference Local file or source How it relates to Qorx
LLMLingua research/papers/llmlingua_2310.05736.pdf Learned prompt compression supports the broader idea that prompts can be shortened while preserving useful information. Qorx core does not bundle LLMLingua.
LongLLMLingua research/papers/longllmlingua_2310.06839.pdf Long-context prompt compression and budget control. Qorx uses deterministic budgeted packing and extractive squeeze in the portable core.
Active Context Compression research/papers/active_context_compression_2601.07190.pdf Active context pruning supports the idea that not all context should remain visible all the time.
Gist Tokens research/papers/gist_tokens_2304.08467.pdf Soft-token memory is a model-side technique. Qorx treats this as adapter/future scope because vendor CLIs cannot consume arbitrary learned gist tokens.
Experience Compression Spectrum research/papers/experience_compression_spectrum_2604.15877.pdf Useful for thinking about compression levels and provenance loss. Qorx keeps exact local fallback instead of relying only on summaries.

Repository And Code Context

Reference Local file or source How it relates to Qorx
ReACC research/papers/reacc.pdf / ACL 2022 source Retrieval-augmented code completion supports bringing related code into the model context. Qorx supplies local code quarks to downstream agents.
Codebase-Memory research/papers/codebase_memory_2603.27277.pdf Codebase memory and graph-style context motivate Qorx’s lightweight symbol and relation surfaces.
BM25 and lexical retrieval referenced in the evidence map Exact lexical retrieval remains a strong baseline. Qorx uses deterministic sparse terms plus path and symbol boosts.
Aider repository map official aider docs Repository maps are useful, but Qorx keeps a local quark store and budgeted evidence routes rather than only a static map.

Agent Memory And Long-Horizon Context

Reference Local file or source How it relates to Qorx
AgeMem research/papers/agemem_2601.01885.pdf Explicit memory operations matter for agents. Qorx exposes local memory CRUD and summaries.
AtomMem research/papers/atommem_2601.08323.pdf Atomized memory matches Qorx’s quark-level evidence approach, but Qorx keeps its implementation deterministic.
Titans research/papers/titans_2501.00663.pdf Neural long-term memory is a real research direction. Qorx does not claim a Titans-like learned memory runtime.
TokMem research/papers/tokmem_2510.00444.pdf Token memory research informs future adapter ideas, not the current portable-core claim.
Memory Survey research/papers/memory_survey_2603.07670.pdf Broad survey background for explicit memory design.
Structural Memory research/papers/structural_memory_2412.15266.pdf Supports structured memory and provenance-aware state.

Hierarchical Memory

Reference Local file or source How it relates to Qorx
H-MEM research/papers/hmem_2507.22925.pdf Hierarchical memory supports multi-layer retrieval. Qorx implements deterministic lattice layers.
HiMem research/papers/himem_2601.06377.pdf Long-horizon memory organization. Qorx uses local mementos and raw-quark provenance.
TierMem research/papers/tiermem_2602.17913.pdf Provenance-aware tiered memory is close to Qorx’s lattice and attestation model.
GAM research/papers/gam_2604.12285.pdf Graph-based agentic memory supports the idea of relations across memory nodes. Qorx keeps graph work lightweight in core.

Cache And Reuse

Reference Local file or source How it relates to Qorx
Preble research/papers/preble_2407.00023.pdf Prefix/cache-aware request design. Qorx has cache-plan for stable prefix and dynamic tail separation.
Don’t Break the Cache research/papers/dont_break_cache_2601.06007.pdf Supports careful prompt structure so provider caches remain useful.
Similarity Caching research/papers/similarity_caching_1912.03888.pdf Approximate reuse can save cost but has correctness tradeoffs. Qorx keeps approximate answer replay out of the default path.
GPT Semantic Cache research/papers/gpt_semantic_cache_2411.05276.pdf Semantic caching supports future guarded adapters. Qorx ships exact replay cache first.
RAGCache research/papers/ragcache_2404.12457.pdf Retrieval cache design for RAG workflows.
Cache-Craft research/papers/cache_craft_2502.15734.pdf Chunk cache management for RAG.
Approximate Caching for RAG research/papers/approximate_caching_rag_2503.05530.pdf Approximate reuse is useful but must be measured and guarded.
Domain-Specific Semantic Cache research/papers/domain_specific_semantic_cache_2504.02268.pdf Domain-specific embeddings can improve cache reuse, but Qorx core avoids mandatory embedding runtimes.
vCache research/papers/vcache_2502.03771.pdf Verified semantic prompt caching supports the idea of cache correctness checks.
ContextPilot research/papers/contextpilot_2511.03475.pdf Long-context reuse. Qorx handles reuse through local handles and evidence resolution.
QVCache research/papers/qvcache_2602.02057.pdf Query-aware vector cache ideas inform future cache adapters.

Runtime And KV Scope

Reference Local file or source How it relates to Qorx
TurboQuant research/papers/turboquant_2504.19874.pdf KV/cache compression is a runtime measurement problem. Qorx can emit hints but does not claim realized TurboQuant/vLLM gains without a runtime proof.
vCache and QVCache local cache papers above Useful for guarded cache reuse and query-aware cache design.

Official Provider And Tooling References

Source URL Qorx note
OpenAI prompt caching https://platform.openai.com/docs/guides/prompt-caching Provider-side cache behavior is separate from Qorx local context omission.
Anthropic prompt caching https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching Qorx can help structure stable prefixes, but provider cache hits must be measured upstream.
Gemini context caching https://ai.google.dev/gemini-api/docs/caching/ Same provider-cache measurement note.
Claude Code memory https://docs.anthropic.com/en/docs/claude-code/memory Memory files are useful, but Qorx adds a live local resolver/index path.
Gemini CLI GEMINI.md https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/gemini-md.md Context files are not the same as budgeted local evidence retrieval.
Cursor codebase indexing https://docs.cursor.com/context/codebase-indexing Cursor’s server-backed indexing is a different deployment model. Qorx keeps the core local.
Cursor secure codebase indexing https://cursor.com/blog/secure-codebase-indexing Useful comparison for privacy and indexing models.

Provenance, Signatures, And Storage

Source URL Qorx note
Protocol Buffers https://protobuf.dev/ Qorx uses protobuf-envelope persisted state and a typed context snapshot.
NIST FIPS 204 https://csrc.nist.gov/pubs/fips/204/final Qorx hybrid attestation uses post-quantum signature practice as a reference point.
C2PA Specification https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html Qorx provenance is local metadata, not a full embedded media manifest.
Microsoft kernel-mode signing requirements https://learn.microsoft.com/en-us/windows-hardware/drivers/install/kernel-mode-code-signing-requirements–windows-vista-and-later- Real RAM-drive drivers have OS/runtime requirements. Qorx reports RAM mode separately from portable disk-backed mode.

PaperQA Result Scope

PaperQA has been used here as a research audit path, not as an oracle. The local corpus supports the architecture class. It does not by itself prove Qorx-specific accuracy, latency, cost, or task-success improvement.

The next benchmark that matters is empirical: multiple repositories, repeated agent tasks, routed provider traffic, retrieval-support scoring, latency, cache hit rates, and invoice/provider-token comparison.