Validated Personal Memory for Language Models: A Bounded Study of State Correctness, Answer Reliability, and Processing Cost · Shounak Pattewale

Preprint · Zenodo · November 2026

Validated Personal Memory for Language Models: A Bounded Study of State Correctness, Answer Reliability, and Processing Cost

Designed and evaluated personal-memory middleware for language models, measuring state correctness, answer reliability and processing cost across nine approaches.

Focus Python · SQLite · LLM evaluationPublished 14 November 2026 · v1.0

Shounak Pattewale · DOI: 10.5281/zenodo.24210707 · Creative Commons Attribution 4.0

The problem

How can an assistant maintain changing personal facts, preserve history, respect declined updates and logically forget records? How do correct stored states relate to correct answers and processing cost?

My contribution

Designed structured personal-memory middleware with proposed operations, staged execution and explicit acceptance. Implemented SQLite records, compared history, summarisation and retrieval approaches, and prepared the manuscript and supporting analyses.

Observed messages feed seven baseline contexts or model-proposed operations. Finalised operations save SQLite records and receipts, which feed answer generation and a separate reference-state audit.
Personal-memory architecture

The study separates saved-state validation from answer generation. Seven baseline contexts are compared with selected-record and all-visible-record approaches using shared reader instructions.

  • Authored middleware comparison: nine methods, three model configurations and 1,620 answer outputs.
  • All 486 saved state-transition traces matched their reference states.
  • All-visible records: 60/60 exact answers for each model. Selected records: DeepSeek 60/60, Gemma 60/60 and GLM 57/60.
  • Selected-record memory used 19.0-26.4% more total model tokens than full history when interpretation was included.
  • Separate model-only analysis: 82,972 saved answers across adapted RULER, LongMemEval-S and LoCoMo-QA. These runs did not evaluate the middleware.
Heatmap comparing nine memory methods across DeepSeek, Gemma and GLM. All visible records score 100% for each; selected records score 100%, 100% and 95%, respectively.
Answer accuracy across nine methods

Results from the bounded study, not a general model-accuracy claim. All-visible records reached 60/60 answers in each model configuration; selected records reached 60/60 for DeepSeek and Gemma, and 57/60 for GLM. Answer accuracy is distinct from saved-state correctness.

Scope and limitations

Published preprint, not peer-reviewed. The authored workload is bounded and synthetic; it does not establish general conversational reliability. Memory lives in external structured storage and does not update model weights. The public repository reproduces released score analyses, not the complete middleware or full end-to-end experiments. The earlier acceptance study retains one failed supported case.

Project chronology

Earlier neural-memory experiments preceded this study. Published on 14 November 2026; this is a publication milestone, not the research start date.

PythonSQLiteLLM evaluationStructured memory

Interested in working together?

Get in touch about AI engineering roles, client projects or research.

Shounak Pattewale · London · 2026

YOUR CHOICES

Privacy & cookies

Choose what you share. The website works with optional analytics switched off.

Necessary

Always on

Remembers your cookie choices for 180 days. No tracking.

Helps understand page visits, how visitors find the site and broad device types. Only collected with your permission and kept for up to 90 days. The analytics cookie expires after 30 minutes without another page visit.

No advertising or personal profiles

Analytics does not store your name, email, form contents, precise location or full IP address.

You stay in control

Change your choice here at any time. Switching analytics off stops future tracking and removes its cookie. For access or deletion of previously collected data, use the contact form.

About this policy

This portfolio is operated by Shounak Pattewale. Hosting services may keep security logs. Fonts are served directly from this website. Linked websites have their own privacy policies. Submitting the contact form sends your name, email and message through Web3Forms so Shounak can reply. These details are not included in website analytics.