Effective Context Capacity for OERSI

Measured synthetic result

The 200 benchmark questions contained each target record's unique identifier. The candidate retrieved the target fact in top 5 and included it in the bounded evidence pack in all 200 trials. A recency-only baseline retaining the final 24 records contained 6 sampled targets (3.0%). Two executions reproduced the same deterministic hit counts and integrity results. This evaluates retrieval mechanics, not semantic understanding or model answer quality.

Architecture

1. Persistent archive: stores original UTF-8 text with source labels, document hashes and ingestion metadata.
2. Overlapping chunks: maintains retrieval-sized slices while preserving the original source text.
3. Lexical retrieval: uses BM25-style scoring plus a small explicit-priority factor; no external model or embeddings.
4. Context compiler: selects whole excerpts under an estimated prompt budget, links each excerpt to document/chunk hashes, and reports omitted candidates rather than silently clipping them.
5. Safety and audit: labels retrieved material as untrusted evidence, abstains on empty searches, and verifies stored hashes.

Token-budget caveat: this prototype estimates tokens as ceil(UTF-8 bytes divided by four). That is a heuristic, not a model tokenizer. A real adapter must use the exact tokenizer and reserve space for system prompts, tools and generated output.

Reproduce it

Python 3.10+, standard library only. No network calls, model inference or private user data are used by the benchmark.

python3 -m py_compile context_memory.py benchmark.py test_context_memory.py
python3 -m unittest -v test_context_memory.py
python3 benchmark.py

Limitations and next gate

The fixture uses query-visible identifiers, so the observed 100% hit rate is not a real-world memory score. The benchmark does not evaluate paraphrase, LLM-generated answers, long-term reasoning, model-specific tokenizers, production load, multi-tenant isolation or adversarial source content. The candidate is not connected to the public OERSI inference route.

The next gate is a held-out long-memory QA test on appropriately licensed public datasets, comparing recency truncation, lexical retrieval, hybrid retrieval and verified hierarchical memory at the same model and token budget. Proposed production criteria: at least a 10-percentage-point answer-accuracy gain, no material increase in unsupported claims, source-grounded citations, enforced tenant isolation and zero unauthorized consequential actions.

Scientific grounding: Titans (NeurIPS 2025), InftyThink (ICLR 2026), MemGPT, FlashMemory-DeepSeek-V4, and Stanford HAI AI Index 2026.