Read the whole relationship before you write one line

Why a relationship AI should assemble the person - email, chat, meetings - before it drafts, and the one thing the research told me I had wrong

Balaji
Balaji2026-09-269 min readSupporter

Most AI email tools read the thread. I decided to read the person.

The thin default versus the assembled relationship: reasoning across a history, not a single thread, is worth about thirty points of accuracy.

Before my system drafts a single reply, it assembles everything it can know about the correspondent: the full email relationship swept backwards from today, the chat corpora where that person shows up, and the transcript of every meeting held with them. Three sources, one person, one context. Then, and only then, it writes.

I want to tell you why I built it that way, what the research says about whether that was a good instinct, and the one thing the evidence told me I had badly wrong.

Sources

  1. LongMemEval: assistants drop ~30% reasoning across sessions, and worse dumping full history vs targeted retrieval (Wu et al., ICLR 2025)arxiv.org
  2. LongLaMP: RAG personalization gives ~+14.9% on the email-writing task, far above fine-tuning (Kumar et al.)arxiv.org
  3. Zep/Graphiti: entity-centric bi-temporal memory, invalidate a superseded fact rather than delete it (Rasmussen et al.)arxiv.org
  4. Reciprocal Rank Fusion: merge score-incomparable rankings by rank (Cormack, Clarke, Buttcher, SIGIR 2009)dblp.org
  5. RAGRoute: a learned router cut up to 80.6% of traffic and 52.5% of latency at equal accuracy vs querying every sourcearxiv.org
  6. MMR: pick each next item to maximize relevance minus similarity to what you already have (Carbonell & Goldstein)elastic.co
  7. Knowledge conflicts for LLMs: inter-context conflict when sources disagree is a named problem (survey)arxiv.org
  8. Microsoft 365 Copilot connectors: per-source federated retrieval as a shipping patternlearn.microsoft.com
  9. Azure AI Search agentic retrieval: LLM query planner fans subqueries across knowledge sources, reranks, mergeslearn.microsoft.com
  10. MIT: personalization features make models more agreeable; a condensed profile produced the largest sycophancy jumpnews.mit.edu
  11. Entity resolution errors structure hallucination rather than reduce itlearningfromdata.zingg.ai

The broken telephone in your AI's memory

Why a relationship AI rots while it still sounds fluent, and the rule that tells you when to rebuild

The filesystem is the last unindexed corpus

Naive RAG over a hard drive will bankrupt you - and what to do instead

Fifteen windows, one handoffs folder

How a fleet of AI task windows coordinates like an engineering team, and the bug that hid two briefs for ten days