Original work from our team.
Papers and research notes from our team on how AI agents use context and memory, and what that means for the cost and quality of AI work.
Field reports
What happened when a team put it to work.
Research notes
Shorter pieces on agents, context and memory.
Each note takes one question about how AI agents work in practice and sets out how we think about it. Newest first.
Research note
Sep 22, 2026
10 min read
AGENTS.md Can Only Have 150 Lines in 2026. Here’s What to Do With the Rest
Why about 150 lines is the real budget for agent instructions in 2026, what to keep, and how to move the rest into checks that run.
Research note
Sep 11, 2026
15 min read
What is coding agent observability in 2026? From tokens to tickets
Traces, sessions, tool calls, instruction-file drift and the token-to-ticket ratio: what observability means for coding-agent spend.
Research note
Aug 13, 2026
10 min read
What Is Agent Procedural Memory? A Plain-English Definition
What agent procedural memory is, in plain English: why agents need memory outside the model, what a stored procedure contains, and how it differs from semantic memory.
Research note
Jul 29, 2026
13 min read
Too Many CLAUDE.md and Skill Files? 4 Fixes for Agent Memory Bloat
When an agent’s instruction and skill files outgrow its context window, there are four fixes: consolidate, restructure, split rules from procedures, or retrieve instead of read.
Research note
May 25, 2026
5 min read
3 Levers for Vertical AI (or Departmental AI)
Foundation models are commoditizing fast, yet vertical AI holds up, because the moat was never the model. Three levers that make vertical and departmental AI defensible.
Research note
May 1, 2026
9 min read
How X-Mem Thinks: A Mental Model for Belief-Aware Memory
Memory isn’t a search problem. It’s a state-management problem: how X-Mem tracks what an agent knows, what it used to know, and why the two can disagree.
Research note
Apr 28, 2026
11 min read
Every Tool Solving the AI Memory Problem Is Solving a Different Problem
Mem0, Zep, GBrain, Karpathy’s wiki, Hindsight: what each one actually does, and the gap they all leave open.
Research note
Apr 28, 2026
9 min read
Your AI Agent Doesn’t Have Memory. It Has a List.
Most AI agents don’t have memory. They have a list. Why context degradation breaks long-running agents, and what belief revision fixes.
Research note
Apr 20, 2026
9 min read
Why Your Vector Database Can Read Your Data And What To Do About It
Why a typical vector database can read what it stores, and how nearest-neighbor search on encrypted data keeps vectors, queries and documents from the server.
Research note
Mar 12, 2026
10 min read
Minimum Viable Context: What Your AI Agent Actually Needs to Know
Your agent doesn’t need total recall. It needs to know what it can’t afford to forget, because the most dangerous mistakes come from context it never thinks to ask for.
Research note
Feb 26, 2026
9 min read
AI Persona Agents: What Happens When Your AI Memory Can Think for You
Your AI already knows how you think, but that understanding is scattered across every tool you use. What happens when it all connects?
Research note
Feb 20, 2026
8 min read
AI Agent Handoff: Why Context Breaks & How to Fix It
Agent handoffs often lose decisions, reasoning and evidence. How structured memory makes multi-agent workflows reliable.
Research note
Feb 17, 2026
12 min read
Why AI Agents Can’t Remember: The Case for the Context Bus
Storage and retrieval are solved, but context still dies when work moves between tools, agents or teammates. The case for a context bus.
Research note
Feb 17, 2026
10 min read
Forward Deployed AI Agents: Why Context Should Scale Even When Engineers Don’t
Forward deployed engineers scale expertise by hand. AI agents can scale it automatically, but only if context moves with them.
Research note
Feb 9, 2026
9 min read
AI Vendor Lock-In: Who Really Owns Your Competitive Advantage?
Your AI vendor holds more than your data. Who owns the advantage built up in it, and how to break the cycle.
Research note
Feb 2, 2026
9 min read
AI Agent Identity: Beyond Authentication
Authentication isn’t enough in the AI era. Why agents need a context login, not just an identity.
Research note
Jan 8, 2026
7 min read
Beyond RAG: Why AI Agents Need Long-Term Memory, Not Retrieval
Retrieval is good at reading a library, but it can’t write an autobiography. Why agents need long-term memory management, not just retrieval.
See what this looks like with your numbers.
We’ll walk through your AI spend, adoption and controls against what leading teams do.