The Agent Context Thread: An Index
Six wal.sh notes on what an agent knows, where that knowledge lives, and who pays for the window

A loose thread of wal.sh notes circles the same question from different angles: what does an agent know, where does that knowledge live, and who pays for the window? No single note on the site owns the framing; six do, between them. This page is an index to those six, in the order a new reader would most profitably take them. Other notes referenced individually (the survey of code search and code graph MCP servers, for instance) cite this thread without joining it; the line is whether the note's primary subject is the context contract itself.

• Agent Memory Architectures: JITIR Against the Field — the deepest cut. Decomposes "memory" into four distinct artifacts (conversation log, working scratchpad, learned facts, reference material) and maps Cloudflare Sessions, MemGPT/Letta, Zep/Graphiti, Mem0, A-MEM, LangMem, and JITIR against the decomposition. The reference vocabulary for the rest of the thread.
• Agent Memory as Institutional Knowledge — reframes the same question for tech leadership. BAML, SageOx, and ThoughtWorks' context graph converge on one move: treat the agent's context like code (versioned, diffable, testable). The bridge from "memory architecture" to "what an organization remembers."
• Claude Code Features: 2026 Q2 — the harness side. 1M context window, Fast mode, hooks, skills, MCP, subagents, worktrees. Reads the agent's context budget as a product surface — where the model ends and the harness begins.
• CLI Coding Agents: 2026 Q2 Comparison — the same harness question across six terminal agents (Claude Code, Copilot CLI, Gemini CLI, Codex CLI, Kiro, OpenCode). Architectural-bet comparison: memory, sandbox, MCP, hooks — the contracts each one offers between the operator, the agent, and its tools.
• 2026 Q2 Skills: Context, Caching, and the Sandbox Lens — skills as the authoring-time half of agent context. Reads SKILL.md through a sandbox lens (compute, fs-read, fs-write, egress, credential) and shows how caching + progressive disclosure make skill composition economically viable.
• Who Does What, With Agents: A CTO's Read of Team Topologies — the synthesis end of the thread. Wulveryck's "systemic context, guardrails, tooling" collapses at small scale into cached system boundaries, each backed by one of the notes above. The thread's outline at organization scale.

The memory notes (1-2) define what context is as a data artifact. The harness notes (3-4) define what the runtime spends to keep it loaded. The skills note (5) defines what the author designs in to keep it small and safe. The team- topologies note (6) reads the whole stack as an org-design problem. Other wal.sh notes are downstream consumers: the code-search and code-graph MCP survey, the deployment-systems notes, the annotation-systems notes, the sandbox-systems notes. They cite this thread; they are not part of it.

A note joins the thread by making the agent-context contract its primary subject, not by mentioning it in passing. Adding one means appending a bullet to 2 above with a one-line annotation, and re-checking that 3 still describes the shape. A note leaves the thread by being subsumed into a successor; in that case keep the bullet and add a (superseded by ...) marker rather than deleting the entry, so inbound links from outside the thread keep resolving.