kinoTou

experiment log

Memory-per-Byte Probe: S vs H under Trace Caps — 2026-06-27

File
2026-06-27_memory_per_byte_sh.md
Size
2.4 KB
SHA-256
00378c636b03d42c…
Download raw file ↓Back to research →
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.

Memory-per-Byte Probe: S vs H under Trace Caps — 2026-06-27

Question

Does architecture H save memory or improve recall under a trace cap, compared to scalar (S)?

Method

Added memory_bytes (pickle size of hippocampus) and --max-traces pruning to multi_fact_stressor.py. Ran real-driver S vs H auto-consolidate with --hybrid-cap 0 and max-traces=1 or 2, with and without reserved-position prefix.

Results

max_traces prefix mode co_recall memory_bytes consolidation_strength
1 no scalar 0/0 10,443 10.0
1 no hybrid 0/0 10,444 36.0
1 yes scalar 1/1 10,443 10.0
1 yes hybrid 1/1 10,444 36.0
2 no scalar 0/0 19,822 10.0
2 no hybrid 0/0 19,822 36.0
2 yes scalar not run — —
2 yes hybrid not run — —

Interpretation

  • Hybrid's ~3.6x higher consolidation strength does not reduce memory_bytes or improve co_recall under trace caps.
  • Prefix still determines exact-token recall.
  • The extra consolidation strength appears to be a latent diagnostic, not a lever that improves measured behavior in these probes.

Implication for architecture

The ingestion scaffold and hybrid modulation are instrumented, validated, and produce the expected mechanical signal. But the S vs H comparison does not favor H on memory-per-byte or exact recall. The architecture H "win" would require either:

  1. A metric that benefits from stronger consolidated traces (e.g. domain-level recall, paraphrase recall), or
  2. A memory model where stronger traces can replace multiple weaker traces (compression), which the current hippocampus/pruning does not implement.

Open questions

  1. What metric does uncapped hybrid consolidation improve?
  2. Can the hippocampus consolidate high-drift chunks more aggressively under hybrid mode, reducing raw trace count?
  3. Should the consolidation-strength cap remain 10.0 as a guardrail, or is it an artificial ceiling that hides useful behavior at higher values?

Artifacts

  • src/oczy/experiments/multi_fact_stressor.py
  • src/oczy/experiments/tests/test_multi_fact_stressor.py

Commits

  • 1c6572a — Add memory_bytes and max_traces.
  • e056161 — Update SUMMARY.md with run #94 result.

Run

Run #94: benchmark code_qa_accuracy=1.0, fast suite 304 passed.