diff --git a/hindsight-docs/blog/2026-03-19-nemoclaw-memory.md b/hindsight-docs/blog/2026-03-19-nemoclaw-memory.md
index a2b9eba7..c3bdd9dc 100644
--- a/hindsight-docs/blog/2026-03-19-nemoclaw-memory.md
+++ b/hindsight-docs/blog/2026-03-19-nemoclaw-memory.md
@@ -8,6 +8,8 @@ image: /img/blog/2026-03-19/nemoclaw-memory.png
 hide_table_of_contents: true
 ---
 
+![Give NemoClaw the Best Agent Memory Available In One Command](/img/blog/2026-03-19/nemoclaw-memory.png)
+
 ## TL;DR
 
 - [NemoClaw](https://nemoclaw.ai) sandboxes isolate AI agents — controlled filesystem, processes, and network. That isolation makes persistent memory harder.
diff --git a/hindsight-docs/blog/2026-03-23-agent-memory-benchmark.mdx b/hindsight-docs/blog/2026-03-23-agent-memory-benchmark.mdx
index 5c959676..11e5f11d 100644
--- a/hindsight-docs/blog/2026-03-23-agent-memory-benchmark.mdx
+++ b/hindsight-docs/blog/2026-03-23-agent-memory-benchmark.mdx
@@ -7,6 +7,8 @@ image: /img/blog/amb/explorer-0.png
 hide_table_of_contents: true
 ---
 
+![Agent Memory Benchmark: A Manifesto](/img/blog/amb/explorer-0.png)
+
 import ImageCarousel from '@site/src/components/ImageCarousel';
 
 We're building Hindsight to be the best AI memory system in the world. "Best" doesn't mean winning every benchmark. It means building something that genuinely works — and being honest about where it does and doesn't.
@@ -15,10 +17,6 @@ That's why we built Agent Memory Benchmark (AMB), and why we're making it fully
 
 <!-- truncate -->
 
-<ImageCarousel images={[
-  { src: '/img/blog/amb/landing.png', alt: 'Agent Memory Benchmark landing page' },
-]} />
-
 ## "Best" is more than accuracy
 
 When we say we want Hindsight to be the best AI memory system, we're not talking about a leaderboard position. We're talking about a complete system that performs across dimensions that actually matter in production:
diff --git a/hindsight-docs/blog/2026-03-23-claude-code-telegram.md b/hindsight-docs/blog/2026-03-23-claude-code-telegram.md
index 31c70572..1b3250b2 100644
--- a/hindsight-docs/blog/2026-03-23-claude-code-telegram.md
+++ b/hindsight-docs/blog/2026-03-23-claude-code-telegram.md
@@ -7,6 +7,8 @@ image: /img/blog/claude-code-telegram.png
 hide\_table\_of\_contents: true
 ---
 
+![OpenClaude: Build a Claude Code Agent with Long-Term Memory — and Take It Everywhere](/img/blog/claude-code-telegram.png)
+
 Anthropic just launched **Channels** for Claude Code: Claude Code sessions connected to messaging platforms. This means Claude Code can now operate as a fully autonomous agent, reachable also from your phone, always running against your codebase (and not just it).
 
 Claude Code has a built-in memory system based on markdown files (`CLAUDE.md`, auto-memory), and it works very well for static preferences and project instructions. But it wasn't designed for conversational memory — it doesn't extract facts from your discussions, doesn't recall relevant context by semantic similarity, and doesn't build up structured knowledge over time. Close the session, and some richness and depth of what you discussed is gone.