Overview
THALIA is an executable research harness for long-context memory systems. It combines typed stage contracts, inspectable context selection, evidence-preserving episodic state, lexical-first retrieval, and bounded compiler search with source-bound evaluation and reproducible artifact checks. The deterministic results are finite diagnostic evidence; model-quality, integrity-advantage, and production-readiness claims remain separately gated.
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- On the fixed 6-example diagnostic set the default configuration reaches composite 0.621 (token-F1 0.477, citation-support proxy 0.830, efficiency 0.854).
- The 8-candidate compiler search ties on this easy set, so configuration improvement is unobserved; its main output is an auditable trace with a tie rule.
- Replacing the extractive Reasoner with gemma3:4b changed mean token-F1 from 0.587 to 0.540 in the 24-example comparison, with an interval spanning zero.
- In the 40-question bottleneck slice, ranking recall is 1.00 but Inspector recall is 0.60; an adaptive context policy raised neural token-F1 from 0.086 to 0.171.
- The author reports operational provenance as the clearest advantage, with equality on poisoning and no general answer-quality advantage.
Methods
- Five-stage typed pipeline: Inspector, Retriever, Reasoner, Memory Gate, Compiler — Implements a long-context memory harness where each stage has a typed input/output contract, runnable implementation, evidence trace and declared scope.
- Offline lexical/hashing retrieval fused by weighted reciprocal-rank fusion — The deterministic core uses BM25/grep, hashing-based retrieval, weighted RRF, real SQLite state and bounded MIPRO/GEPA-style configuration search without network calls.
- Episodic-first memory: raw episodes appended before gated consolidation — The Memory Gate appends each raw turn to state before evaluating a derived memory delta, preserving source material for auditing.
- Tiered evaluation on LongMemEval-style, local, and LongMemEval_S slices — Evaluates a fixed 6-example diagnostic set, a 24-example local comparison with gemma3:4b, an external LongMemEval_S transfer slice, and a 40-question bottleneck slice.
- Injection and poisoning probes of cited excerpts — Recorded injection probes test whether cited excerpts contain a supplied gold value and measure a poisoning effect, as negative controls.
Summary sources: Paper metadata and evidence · Extracted source text.
Document downloads
PDFs are hosted here; supporting sources may use pinned archives.
- thalia-0.1.0.pdf PDF · 4.77 MiB · extracted-text source
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related works
Ranked by recorded methods, topics, contributors and associated software, with contributor roles and domain-only matches labelled. Metadata overlap is not a citation or a scientific equivalence.