Sign InOpen Brain
AI EngineerVideoSource Linked

Scaling Compute on Context — Jack Morris, Engram

Training directly on a private corpus can memorize it without producing useful behavior. Durable context learning likely needs synthetic study tasks and recursive self-improvement, not a single fine-tuning pass.

AI Engineer · Aug 12, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Morris frames the problem as teaching a pretrained model an unstructured private corpus. Direct next-token training on **10,000 financial reports** can drive loss to **0.00001** yet still produce collapsed generations, showing that memorization is not usable knowledge.

Practical Implication

For private code, messages, or documents, distinguish retrieval from learning. Context compression, on-policy distillation, synthetic question generation, and continued pretraining offer different tradeoffs; the desired system should keep generating harder study material as the model improves.

Agent-Ready Context
Morris frames the problem as teaching a pretrained model an unstructured private corpus. Direct next-token training on **10,000 financial reports** can drive loss to **0.00001** yet still produce collapsed generations, showing that memorization is not usable knowledge.

For private code, messages, or documents, distinguish retrieval from learning. Context compression, on-policy distillation, synthetic question generation, and continued pretraining offer different tradeoffs; the desired system should keep generating harder study material as the model improves.

Every described method has limits: context methods require the data to fit, synthetic training can overwrite prior knowledge, and one-pass approaches plateau. Recursive self-improvement is presented as the research target, not a settled recipe.
Connected Context · Feed7 Judgment

This rules out low training loss as evidence that a private corpus has become usable knowledge and turns corpus adaptation into a choice among retrieval, compression, and parameter updates. It reinforces hybrid memory designs and workload-specific evaluation, while warning that consolidation can overwrite prior knowledge and that recursive improvement remains an open target.

UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task StreamsUniMem operationalizes the retrieval-versus-learning distinction by routing novel episodes to retrieval and recurring patterns to parameters, but does not settle overwrite risk.Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist DomainsThe Greek RAG study reinforces the need to evaluate corpus access methods on the actual domain rather than assume learned or dense representations will outperform lexical retrieval.Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State InjectionPRECOG offers an architecture-specific way to reuse corpus context without retraining or repeated ingestion, covering the retrieval side of the tradeoff.Citation Needed: Provenance for LLM-Built Knowledge Graphs — Daniel Chalef, Zep AIProvenance becomes an implementation requirement when compressed, synthesized, or learned corpus knowledge can no longer be traced through simple source pointers.
Context Map
contextdataresearch#context-engineering#retrieval#agent-memory
Uncertainty
Every described method has limits: context methods require the data to fit, synthetic training can overwrite prior knowledge, and one-pass approaches plateau. Recursive self-improvement is presented as the research target, not a settled recipe.