AI memory is moving from retrieval into model adaptation. Engram's Dan Biderman and Jessy Lin argue that the bottleneck is not raw intelligence, but whether models can internalize evolving team context instead of rereading it through giant prompts.
Enterprise AI may split memory by workspace, not collapse it into one universal personal profile. The Engram discussion points toward team-level and product-level memory, with control boundaries between work, personal, and application contexts.
Fine-tuning is being reframed as a live product primitive. Rather than treating pretraining and post-training as one-time phases, Engram describes a world where models keep learning through adapter tuning, supervised fine-tuning, RL, and distillation pipelines.
The economic pitch for learned context is fewer tokens and better behavior on bespoke work. Engram claims deeply learned team context can reduce the need for massive system prompts and repeated retrieval, sometimes by orders of magnitude.
Coding agents are turning intermediate work into collaborative artifacts. Claude Code artifacts make agent sessions shareable as live pages, shifting status updates, PR walkthroughs, incident timelines, and architecture explanations from chat transcripts into persistent team surfaces.
No new X / Twitter builder posts were present in this week's feed.
Training Data: Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin
The Takeaway: Engram's bet is that AI memory will not be solved by stuffing more context into prompts. The more durable path is teaching models to internalize what matters about a team, product, or workflow.
Dan Biderman and Jessy Lin, cofounders of Engram, describe the company as working on memory and continual learning for models that are not just queried against a workspace, but trained to understand it. Their core claim is refreshingly specific: today's bottleneck is "not really raw intelligence," but understanding new and evolving context. In a company, that means learning initiatives, hiring practices, tools, priorities, and tacit ways of working the way a long-tenured employee would.
Their product philosophy is not anti-RAG or anti-tooling. External memory still matters. The sharper distinction is what should stay external and what should be internalized into weights. Engram trains per-team models inside workspaces, using approaches such as adapter fine-tuning, LoRAs, supervised fine-tuning, RL, and on-policy distillation. The goal is to turn documents and interactions into training signal, not just context-window material.
The economic argument is the most concrete part. Jessy Lin says learned context can avoid "monstrous system prompts" and, in some cases, lead to "100x fewer tokens" because organization-specific knowledge is often implicit across many documents and conversations. Instead of asking a frontier model to scan 100,000 tokens to infer a company's operating style, a tuned model may answer from internalized context in a much smaller budget.
There are real constraints. Engram needs white-box access to weights, which makes open-source models easiest and closed frontier models possible only through partnerships. The founders also resist the idea of one universal memory wallet: personal, work, and product memories may need to stay separated so users can control what a model remembers where.
https://www.youtube.com/watch?v=aiR7F4jqjXY
Claude Blog: Claude Code now supports artifacts
Claude Code artifacts turn an agent session into a live, shareable web page: PR walkthroughs, system explainers, dashboards, release checklists, incident pages, and other visual outputs built from the session's context. The important shift is not just prettier output. It is that Claude Code can use the codebase, connectors, and conversation to create a page teammates can inspect without replaying the whole chat.
The strongest use case is debugging and coordination. Anthropic describes incident work where Claude Code publishes a timeline, suspect commits, and an error-rate chart, then republishes the same artifact as the investigation progresses. That changes the collaboration surface: instead of asking one engineer to verbally summarize what the agent found, the team can look at the same evolving artifact.
Artifacts are private to the author by default, can be shared with authenticated teammates or the organization, and include version history. Admins get organization-level controls, retention policies, and compliance visibility. The feature is in beta for Claude Team and Enterprise users from the Claude Code CLI and desktop app, with pages viewable in a browser.
https://claude.com/blog/artifacts-in-claude-code
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders