fXIGinM Wednesday, 18 November 2026
Technology

Why AI Needs Memory It Can Trust

Memory Context Provenance and Persistent Understanding By Sachin Dev Duggal, Founder and CEO, SeKondBrain AI/SI Imagine working with a brilliant colleague who needs you…

AI memory and data provenance concept illustrating trusted context sharing across AI assistants by SeKondBrain AI.

Memory Context Provenance and Persistent Understanding

By Sachin Dev Duggal, Founder and CEO, SeKondBrain AI/SI

Imagine working with a brilliant colleague who needs you to reconstruct your project every morning. Yesterday’s decisions, the options you rejected, the reason an apparently sensible approach will not work. We would recognise that as an expensive way to collaborate. Yet we often accept it when working with AI.

We repeat briefings. Corrections disappear. Moving between AIs means rebuilding the background. An assumption returns as an apparent fact. As models improve, I suspect the ability to carry useful context forward could become more valuable than another increase in model size or context capacity.

More context needs better organisation

A model’s parameters encode patterns learned during training. Its input context contains the request, instructions and perhaps earlier messages or retrieved material. An ordinary conversation does not, by itself, rewrite those parameters. Persistent memory can instead live outside the model, with relevant information retrieved into later requests.

A larger context window lets a model receive more information at once. It does not decide which version is authoritative or which decision still applies. Think of a library: expanding the reading room lets us spread out more books, but we still need to find the relevant edition.

The 2023 study Lost in the Middle found that information placement affected performance in its tested models and tasks. It reminds us to assess how context is used, alongside how much is accepted.

From evidence to reusable concepts

Our arXiv paper, Atomic Units of X: The Compression Layer of Intelligence, which I co-authored with Pradyumna Swarnalatha Ramanna and Alexandros Vassiliades, proposes reusable conceptual structures that connect related evidence while preserving its sources.

The latest version evaluates evidence-to-concept consolidation across HBase, Hadoop and Maven software corpora. Under its deduplicated paragraph-level method, HBase and Hadoop exceeded 20 evidence paragraphs per reference concept; Maven did not. Results depend on corpus scale, evidence boundaries and concept identity. These are consolidation measurements, not equivalent token savings or proof of a complete reasoning architecture.

The practical idea is accessible. A project’s registration requirement may appear across documents and conversations. An identifiable concept can connect those records, including why open registration was replaced by an invitation-only pilot. The model can navigate that structure without treating every mention as an independent requirement.

Compaction must preserve what matters

Conversation compaction addresses another question: how much history must the model read to continue? A digest can retain decisions and recent exchanges, with routes back to supporting evidence.

Our separate internal replay of 1,314 sessions found that estimated savings depended strongly on conversation length. Short interactions could cost more after memory overhead; longer conversations drove most savings. The replay omitted recalled-memory payloads and reconstructed digests. Its estimates are upper bounds, not production billing results; provider caching also changes the comparison.

The principle is to measure efficiency against the actual workload. Fewer tokens are useful only if the necessary correction, constraint and qualification survive.

Provenance makes memory accountable

A transcript records what was said. Useful memory distinguishes a suggestion from a decision and an assumption from verified evidence. Provenance records where information came from, when it applied and how it changed.

Without that history, an AI-generated suggestion can be stored and later presented as an agreed requirement. Repetition gives it apparent authority. Source links should lead to supporting passages, with inferences identified separately. When authoritative sources conflict, the system should expose the disagreement. The newest document may be an unapproved draft.

Persistent understanding should be judged through behaviour: does a correction survive the next session, and does the system recognise when the circumstances behind a decision have changed?

India and the memory passport

India makes continuity a practical question across languages, channels and domestic AI systems. A person might explain a requirement in Hindi, review a document in English and return through voice. The intent and evidence should survive those changes.

BHASHINI’s integration with e-Shram, announced in January 2025, extended the portal to all 22 Scheduled Languages. That demonstrates wider linguistic access. Preserving context across permitted interactions is a further challenge.

Your accumulated context is your memory, and it should be your passport to the AIs you choose.

Imagine carrying relevant ChatGPT history into a conversation with a Sarvam-powered assistant or an Indian personal AI agent. Your decisions and preferences should help the conversation continue. Trying a domestic AI should not require rebuilding your working life from a blank prompt.

This requires compatible interfaces that carry meaning, provenance and permissions together. Users must choose what travels. Indian organisations should also know where context is stored and processed, and which providers receive it. Domestic model choice and memory governance belong in the same discussion.

Enterprise value and three risks

For enterprises, continuity should mean less repeated briefing and less rework from forgotten requirements. A product assistant, coding assistant and support assistant could share the approved pilot decision while receiving only the context each needs. Personalisation should preserve explicit preferences that users can inspect and correct. Measure the workflow improvement, rather than treating stored-memory volume as success.

The same continuity can preserve mistakes. Three risks deserve attention.

False information becoming durable memory

An invented exception can enter a summary and influence later work. Malicious instructions in external material can also contaminate context. OWASP identifies prompt injection and retrieval poisoning as risks. Organisations should separate evidence from instructions and require traceable support before promoting generated claims into approved decisions. Consequential memory changes may need human review.

Private context crossing permission boundaries

A support assistant may need the pilot’s registration rule, but not confidential customer discussions. Enforce access controls before protected context reaches the model. Summaries and derived memories need protection too: revoking access to a document is ineffective if its sensitive contents remain retrievable through a summary. A prompt asking the model to keep secrets is insufficient.

Outdated decisions shaping new actions

If the pilot ends, faithfully repeating its old restriction becomes an operational error. Record when information applies, who owns the decision and what replaces it. Retire obsolete instructions from active use while retaining the historical evidence where appropriate.

Test continuity before trusting it

Start with a bounded workflow. Correct a requirement, then revisit it in another session and through another permitted AI. Check whether the correction and its supporting evidence survive. Repeat with a user lacking access, revoke permissions, introduce conflicting instructions in an untrusted document and replace an old decision.

These tests examine retrieval, storage and access enforcement around the model. Where AI takes consequential action, retain appropriate approval controls. Remembering a past decision does not authorise a new action.

Model capability remains essential. But when products access comparable models, how they maintain and use context becomes a plausible source of advantage. Providers should earn loyalty through usefulness, while users remain free to carry their memory elsewhere.

That belief sits behind our work at SeKondBrain on Kemory: a shared memory layer through which useful context can accompany people across their chosen AIs, including domestic models and personal agents.

Our brilliant colleague becomes more useful when we can resume the work together. The question I would ask of AI is equally ordinary: when I return tomorrow, can we continue from where we left off?

Editorial notes for SeKondBrain

Internal only remove these notes before submission

Revised 14 October 2026 for TechNadu with India, the memory passport framing, Atomic Units of X and compaction analysis. Complete article: 1,206 words including headline, subtitle, byline and subheads, within Sachin’s 1,350-word maximum. These internal notes are separate. No submission was made.

Status check

The status check below dates from 13 October. No newer draft or submission was found then; private email and inaccessible sources were not ruled out. This revision changes the article, not the task status.

TechNadu S9N-10557 was Planned, due 14 October, minimum 800 words. CIO&Leader S9N-10552 was Backlog, approval pending; revised 28 October conflicts with the original 14 October brief. Confirm the operative deadline.

Distinct treatment and exclusivity

TechNadu treatment: continuity, India and portable context, research boundaries and memory security. No CIO&Leader draft was found for comparison. Give any separate CIO&Leader piece different wording and examples, focused on executive governance and procurement.

Exclusivity terms were absent from the retrieved briefs. Before submission, confirm first-publication rights, exclusivity period, simultaneous-submission restrictions and later adaptation rights. Disclose the separate same-topic opportunity. Do not submit the same article to both outlets or cross-post before these terms are resolved.

Evidence and author disclosure

The byline and Kemory paragraph disclose the commercial connection. Competitive advantage and memory portability are arguments and architectural goals. The ChatGPT-to-Sarvam example does not claim a verified live integration. Examples are hypothetical, not customer incidents.

Lost in the Middle 2023 concerns its tested models/tasks. Retrieval Augmented Generation 2020 supports external retrieval, not guaranteed correctness.

OWASP Prompt Injection · Vector and Embedding Weaknesses · Sensitive Information Disclosure. These support the security risks and layered controls; the workflow tests are the author’s recommendations.

Atomic Units of X arXiv 2607 12634 v3 was checked against the 27 August 2026 PDF. Its software-corpus consolidation results are not equivalent to token savings or proof of the full architecture. Older 46.2% and 47.5% figures on the company research webpage were not used.

Compaction source: supplied token_analysis (1).pdf, 26 September 2026, replay at commit 528023e. The 72.2% input / 61.3% overhead-adjusted headlines were omitted because recalled payloads are absent, digests reconstructed and caching modelled. No customer-billing, competitor-ranking or India-specific inference is made.

PIB e Shram and BHASHINI 7 January 2025 supports linguistic access only. Sarvam sovereign LLM announcement supports the domestic-model example, not memory integration or blanket residency.

News Outreach

News and explainers from the News Outreach editorial team.