Memory Is Having a Moment. Tech & AI Weekly by Eli, August 14 2026
I took Friday and the weekend fully offline, no laptop, no news. Back Monday with a full inbox and two weeks of stories to sort through. Here’s what actually mattered.
Memory showed up everywhere: in a benchmark that went from 35% to 98% by fixing context, in an Anthropic technique that lets agents consolidate what they’ve learned, and in three posts I published on exactly that problem. Coincidence, but a useful one.
AI & GenAI
-
Context engineering: task correctness went from 35% to 98%. A presentation on why stuffing prompts doesn’t scale: lazy-loaded skills, versioned context artifacts, and externalized memory banks outperform a 100k-token context window. The numbers are from a real benchmark, not a thought experiment. Read more
-
Meta open-sourced Muse Glimmer 30B for local devices. Apache 2.0, quantized to under 20GB, runs on a single consumer GPU. Tool use, multimodal input, 100+ languages, no cloud required. On-device agentic AI just got a lot more credible. Read more
-
Vercel launched Agent Plugins 1.0.0, backed by OpenAI, AWS, Cursor, GitHub and Microsoft. A portable, vendor-neutral format for packaging reusable AI agent components (skills + MCP servers). Five competing companies agreeing on a format is unusual enough to pay attention to. Read more
-
Anthropic’s “dream memory”: agents consolidate memory during idle cycles. Presented at AI DevCon London by Anthropic staff. Interesting technique, but researchers are already flagging a real risk: if an agent builds bad habits, consolidation can turn them into permanent policy. Read more
Cloud & DevOps
-
GitHub to maintainers: “Your contributors are AI-first. Is your project?” Practical guide: put
AGENTS.mdbeside the code it governs, enforce PR templates via CI as a hard wall, use CLAs as human detectors. Structural enforcement, not policy documents. Read more -
Study: coding agents ignore open source guidelines even when loaded in context. Peking University, 4 frontier models, 106 issues across 49 repos. Rewriting your CONTRIBUTING.md won’t fix it. Read more
-
Databricks acquired Electric to give each AI agent its own Postgres. Electric uses WebAssembly to spin up lightweight, isolated Postgres instances without heavy infrastructure. The bet: agents need their own data layer, not shared state. Read more
✍️ From me lately
-
Stop Your AI Agent Forgetting User Preferences: Key-Value Memory. Conversation history feels like memory but dies when the process restarts. I show how to wire
agent.state+ session managers (File and S3) to persist structured preferences across restarts. No vector database, no embeddings required. Read more -
AI Agent Memory: Add Semantic Search Without a Vector Database. Key-value memory breaks when the query shares no words with what’s stored. The fix: embed each memory at write time, retrieve by cosine similarity. I compare FAISS (in-process, ~0.09ms) vs Amazon S3 Vectors (~195ms) and show how to let the agent self-route between KV and semantic memory via tool docstrings. No routing logic, no prompt engineering. Read more
-
Build AI Agent Harnesses in a Free Sandbox Workshop (AWS Builder Center). Practice Strands Agents patterns in a real AWS environment, free, no account setup needed. Follows the same progression as Morgan Willis’s 14-video course and ends with a deployed agent in the cloud. Available in English, Spanish, and Portuguese. Read more
-
Running an Event Stand with Strands Agents? Here Is Your Demo Kit (AWS Builder Center). A split-screen web app for event stands: visitors type into the left panel, the right panel shows every reasoning step, tool call, blocked action, and token spent in real time. 11 live demos, deploys with AWS credits,
cdk destroycleans up after the event. Read more -
Up next: I’m finishing the memory series and then moving to a caching piece. I built a semantic cache for Strands Agents using Valkey, measured the results, and they’re worth writing about. Stay tuned.
📸 Where I was
Last week I was fully offline: personal trip, no laptop. First real disconnect in a while. I’m back now, inbox at capacity, and this edition is part of the recovery.
I hope you enjoy this as much as I enjoy putting it together every week. It’s my little weekly de-stress moment. Let’s stay informed together and learn something new along the way.
See you next Friday,
Eli
Dev.to · LinkedIn · GitHub · Twitter/X · Instagram · YouTube