Hindsight 0.10.1 is out. A new reranker that takes recall@1 from 0.800 to 0.950 while running 4x faster, plus a bank you can finally pick up and move: clone it, export it, rename it in place. Highlights below 🧵

Knowledge Graphs vs. Vector Search for Agent Memory. Almost every "add memory to your agent" guide reaches for a vector database on the first line. A smaller, louder camp says the opposite: real memory is a knowledge graph, and embeddings are a toy. Both camps are half right, and the half they are missing is the half that breaks in production. The honest answer is that vector search and knowledge graphs solve different halves of the memory problem. If you pick one and ignore the other, you will pass the demo and fail the first hard question a real user asks. Here is where each one wins, where each one breaks, and why agent memory in particular needs both. Know more: linkedin.com/pulse/knowledge…

Your Agent Can Now Manage Its Own Knowledge Base Over MCP. A Hindsight knowledge base is the readable, self-healing wiki sitting on top of your agent's memory: pages that distill what the agent knows into something a human can skim, and that rebuild themselves as the underlying memory changes. Until now there was an odd asymmetry. An agent could retain and recall over MCP all day, but to browse or maintain that wiki it had to drop down to a separate HTTP path. The memory was native; the knowledge base was a second integration. Hindsight 0.9.2 closes that gap. The knowledge base is now a native MCP surface: seven agent-facing operations registered as tools, on the same connection the agent already uses to remember. Know more: linkedin.com/pulse/your-agen…

Hindsight now includes @typesafeai Jev reranking capabilities!

Cross-Encoder Reranking: The Last Stage of Agent Memory Recall. Retrieval is the part everyone talks about. Embeddings, BM25, graph traversal, time windows: pick your arms, fuse the rankings, ship it. But fusion doesn't tell you what's relevant. It tells you what several independent systems each thought was plausible, blended into one list. The stage that decides what your agent actually reads is the one that runs last, and it's the one that gets the least attention. This is a look inside that stage in Hindsight — the cross-encoder reranker. How it scores, why the secondary signals are multiplicative instead of additive, the subtle failure mode that turns a reranker into a date sort, and how to configure it so an unreachable model degrades instead of taking recall down with it. Know more: linkedin.com/pulse/cross-enc…

Migrating agent memory carries the old system's duplicate conclusions with it. Worse: imported facts arrive marked as already consolidated, so your bank never rebuilds the beliefs itself. Bring the source material. Leave the conclusions.

Great Coding Agents Don't Just Reason. They Remember. Our New Plugin Makes That Standard. Hindsight 0.9.0 ships two things we've been building in parallel for months: Knowledge Pages — living documents your memory bank writes and maintains about itself — and a unified coding-agents plugin that brings long-term project memory to ten coding agents — Claude Code, Codex CLI, Cursor CLI, opencode, GitHub Copilot CLI, Cline, Kilo, Grok Build, Antigravity, and Devin — with one command and zero configuration. Neither looks like what we set out to build. This post is the story of how we got here, because the detours are the product: a benchmark we had to invent, an architecture of ours it killed, and a wiki that learned to heal itself. Know more: lnkd.in/gU9YvzhY