Ricord / Blog
Research & engineering notes
Benchmark runs, architecture decisions, and what we're learning as we build the memory layer for AI agents.
- Guide6 min read
ChatGPT Has Memory Now — But It's Locked In. Here's How to Make It Travel (2026)
ChatGPT remembers you across chats — but only inside ChatGPT. Tell it your stack on Monday and Claude, Cursor, and your own app have no idea on Tuesday. The honest take on ChatGPT's native memory, and how to wire in portable memory you actually own.
Read post - Guide7 min read
The Best AI Memory for Every Coding Editor (2026)
Cursor, Windsurf, Cline, Claude Code, Codex — every AI coding tool is brilliant inside a session and amnesiac across them. Each ships its own half-answer: rules files, a memory bank, session history. How editor memory breaks into three tiers, and how to pick the right one for how you work.
Read post - Engineering11 min read
The Token Economics of Agent Memory: When a Memory Layer Pays for Itself
Most teams pick a memory layer the way they pick a logger — by feature, not by cost. But the deeper you go on agent products, the more memory becomes a unit-economics decision. The math on when paying for a memory layer actually saves money, with a worked example at 1k, 10k, and 100k MAU.
Read post - Engineering10 min read
How to Evaluate a Memory Layer in 1 Hour: A Buyer's Checklist
Picking an AI memory layer for a production agent is a decision you'll live with for years. You don't need a quarter to evaluate — you need an hour spent on the right tests. Eight 5-minute checks that will tell you more than a vendor demo.
Read post - Engineering12 min read
Multi-Tenant AI Memory at 10k Users: A Production Playbook
Memory works great in the demo. Then you have 10,000 paying users, and the same patterns that shipped your MVP become a series of expensive failure modes. The five problems that show up at scale — isolation, noisy neighbors, recall cost, GDPR delete, and per-tenant budget — and what each layer should do about them.
Read post - Security11 min read
Defending Agent Memory: A Layered Model for 2026
Memory poisoning is the active threat. The defense isn't one trick — it's six layers, each cheap in isolation, none sufficient alone. Here's the model production teams ship.
Read post - Engineering9 min read
Adding Cross-Thread Memory to LangGraph: A Worked Example
LangGraph's Checkpointer handles thread state. LangGraph's Store is a primitive, not a product. Here's what it actually looks like to add cross-thread, multi-tenant, conflict-resolving memory — both the build-it-yourself version and the wire-in-Ricord version, side by side.
Read post - Engineering11 min read
How AI Agents Actually Remember (An Architecture Field Guide)
Cognitive science has four memory categories. LLM agents need all four, but the architecture for each is different. A walk through working / episodic / semantic / procedural memory in production agents — what works, what breaks, where teams land in 2026.
Read post - Engineering10 min read
Open-Source vs Hosted Memory Layers: When Each Wins
Every AI memory layer now ships in two flavors — open-source self-host or hosted SaaS. The right answer depends on five axes, not on ideology. Here's the framework, slot-by-slot.
Read post - Engineering8 min read
MCP Memory Server: A Developer's Guide
What an MCP memory server actually does, the minimum-viable version (three tools), and what separates a toy from a production one. Plus the 60-second install path.
Read post - Engineering7 min read
How to Make Claude Code Remember Across Sessions
Claude Code is brilliant within a session and amnesiac between them. Three patterns that fix it — including an MCP-native memory layer that installs in 60 seconds.
Read post - Security9 min read
Memory Poisoning: The AI Attack Vector Nobody Was Watching For
An active threat to enterprise AI agents. What it is, why most memory APIs fail, and what to evaluate.
Read post - Engineering7 min read
Why Your AI Agent Forgets: Five Problems Nobody Talks About
The five hardest memory problems in production agents — and what actually works.
Read post