Blog
Insights & research on AI memory.
Deep dives on persistent memory for agents — the pain of forgetting, the difference between memory and RAG, and the case for memory that predicts, not just recalls.
We Beat Mem0 and Zep on Their Own Benchmarks — at 13× Fewer Tokens
We scored our memory engine on LOCOMO using each competitor's own judge — Mem0's partial-credit rule, Zep's same-topic rule. We win overall against both, and feed the model 522 tokens per query instead of ~7,000. Every number is reproducible.
The AI Memory Landscape in 2026: Where Brains Fit
We mapped the field — Mem0, Zep, MemoryLake, Letta, LangMem, Cognee, Supermemory, Memobase. Everyone is racing on recall. Two frontiers sit wide open: forecasting, and a marketplace of installable brains.
We Benchmarked a Brain Honestly: Calibration Beats Accuracy
A brain is grounded memory a model installs instead of stuffing context. On BEAM-100K it matched naive RAG's answer quality on 1/21st the tokens — and knew when it didn't know, at 2.5× the abstention. The honest read, weak spots included.
From AI Memory to an Expert Brain: A Practical Guide
Chat history isn't memory. A bigger context window isn't a brain. Here's what persistent memory actually is — and what turns it into a brain an agent can install.
Mem0 vs MemMesh: A Memory Store vs. an Installable Brain
Mem0 is the category's most-adopted memory layer. MemMesh is built on two things Mem0 doesn't do — calibrated prediction and installable, publishable brains. An honest, sourced comparison.
Zep vs MemMesh: A Temporal Graph vs. a Brain That Forecasts
Zep's temporal knowledge graph is one of the best in the field, and its provenance is genuinely strong. Here's where MemMesh differs — prediction, installable brains, and on-device — and where Zep wins.
MemoryLake vs MemMesh: A Memory Passport vs. an Installable Brain
MemoryLake makes your context portable across every AI, with real provenance and versioning. MemMesh shares that portability goal — and adds prediction and a marketplace of installable brains. An honest comparison.
Why Claude Code Forgets Your Project — and How to Give It a Brain
Claude Code opens each session with a fresh context window and a static CLAUDE.md. Here's why it forgets, and how to install a project brain in one command.
Install a Brain into Claude Code in 2 Minutes (MCP)
A step-by-step guide to installing a MemMesh brain into Claude Code over MCP, so your agent remembers across sessions and predicts what's next.
A Brain, Not a Notebook: Why Recall Alone Isn't Enough
The first wave of AI memory was about remembering. The next wave is a brain that anticipates. Here's the case for prediction as a first-class part of a brain.
The EU AI Act Wants Provenance — What That Means for AI Brains
Transparency and traceability obligations land on systems that remember. Here's what provenance, retention, and erasure mean for a brain your agents install.
Calibrated Confidence: How a Brain Scores Its Predictions Against Reality
A confidence score is only useful if it's honest. Here's how calibration, abstention, and a reconcile loop keep a MemMesh brain's forecasts trustworthy.
Brain vs. RAG vs. Vector DB: When You Need Which
Three things that get lumped together and shouldn't. A clear map of what each does, where each fails, and how a brain sits on top of them.
One Key, Every AI: How MeshKey Is Your Portable Brain
Your brain shouldn't be trapped inside one vendor's chat app. Here's how MeshKey gives you a portable, private brain that unlocks everywhere you work.
From Stateless to Trusted: Giving Financial AI a Brain
A financial advisor that forgets your portfolio every session can't be trusted with it. Here's what a finance brain — grounded memory plus calibrated prediction — changes.