Blog
Engineering deep-dives, release announcements, and community notes from the OpenZync team.
- 2026.11.02What breaks when three write paths disagree about the same fact — and what a reproducible LongMemEval-S run, weak categories included, does and does not show.
OpenZync: Converging Write Paths for Agent Memory — A Technical Report
engineering - 2026.10.10We ran LongMemEval-S end to end — 500 questions, full methodology disclosed. The number is mid-band in 2026's leaderboard, and the reasons the field's numbers don't compare are the real story.
87.1% on LongMemEval-S: The Harness Matters More Than the Score
engineering - 2026.09.23Everyone upgrades the read path. Three September preprints show the write path is where memory actually breaks — and where it gets fixed.
Smarter Retrieval Won't Fix Messy Memory
engineering - 2026.09.22Every memory write and every retrieval routes through autoregressive generation. One September preprint moves the bookkeeping to a cheap control plane — and cuts build time 6.6×.
Your Agent Runs Its Memory on a Frontier LLM
engineering - 2026.09.01Your agent doesn't need a bigger context window — it needs a memory harness that survives fifty turns. Why dumping two million tokens still fails.
Your Agent Loses Its Mind After 50 Turns — And 2M Tokens Won't Save It
engineering - 2026.08.28Everyone teaches chunk — embed — retrieve for agent memory. The products that actually remember you — ChatGPT and Claude — don't ship it at all.
Your RAG Pipeline Is Not How ChatGPT Remembers You
engineering - 2026.08.25Serving cost for agent memory is unpredictable, unbudgeted, and growing. The cheapest-looking architecture — resending everything — is often the most expensive.
Cheap Memory Is the Expensive Option.
engineering - 2026.08.24GDPR erasure and AI Act logging now reach into agent memory stores. Deleting one fact means chasing it through transcripts, embeddings, and model behavior.
The Right to Be Forgotten Meets Agent Memory.
engineering - 2026.08.23The agent stack's protocols just finished consolidating under one neutral foundation. Shared memory isn't in either spec — by design. That gap is about to matter more.
MCP and A2A Have a Foundation Now. Memory Still Has No Standard.
engineering - 2026.08.20Anthropic and EPFL researchers showed ideas spreading between AI agents through their memory files — surviving every wipe. The carrier is the part you never audit.
Your Agent's Memory File Is How Mind Viruses Spread.
engineering - 2026.08.18Twenty-eight percent of the web ships an llms.txt. Ninety-seven percent of those files were never fetched by anyone. What agents actually read — and what to build instead.
Everyone Ships an llms.txt. Almost No Agent Reads It.
engineering - 2026.08.17Context engineering replaced prompt engineering — and it still won't fix your agent's memory. What 9,649 experiments actually show.
Context Engineering Won't Fix Your Agent's Memory.
engineering - 2026.08.16Vector database, knowledge graph, or raw logs for agent memory? A fresh benchmark says search over the unmodified record beats structured memory — and what still needs a graph.
Your Agent Doesn't Need a Knowledge Graph. It Needs a Search Box.
engineering - 2026.08.15False memory isn't a machine bug — humans have it, AI amplifies it, and agents store their own confident lies. The 2026 evidence and what to do about it.
Your Agent Remembers Things That Never Happened.
engineering - 2026.08.14LoCoMo-saturated agents fail realistic multi-session tasks — the 2026 benchmarks exposed the gap between memorization scores and agentic memory.
Your Agent Aced the Benchmark. It Still Doesn't Remember.
engineering - 2026.08.12Tencent open-sourced a team memory hub this week. Sharing is solved; correcting is not — who fixes the wrong fact every agent now inherits?
Team Memory Is Here. Nobody Agrees on How to Fix a Wrong Fact.
engineering - 2026.08.09Prompt injection ends when the session ends; memory poisoning persists — one write into long-term memory steers every future interaction. The 2026 attack surface and the architectural defense.
Poison Once, Exploit Forever: The New Attacks on AI Memory (and How to Defend It)
engineering - 2026.08.08Agent2Agent lets agents talk to each other — but neither A2A nor MCP provides shared memory. The gap, five inadequate patches, and what a third protocol layer needs.
When Agents Talk to Each Other, Who Remembers? A2A, MCP, and the Shared-Memory Gap
engineering - 2026.08.07The Model Context Protocol made AI memory a pluggable feature. A field guide to memory servers in 2026: how they work, the storage-model taxonomy, the security caveats, and how to set one up.
MCP Memory Servers in 2026: A Field Guide to Giving Your AI Assistant a Real Memory
engineering - 2026.08.06A pattern-by-pattern field guide to agent memory — episode ingestion, entity resolution, temporal validity, hybrid retrieval, and community detection — with the tradeoffs that decide which to build.
Building with Graph Memory: A Field Guide to Agent Memory Patterns
engineering - 2026.08.05Mem0 and Zep are the most-installed memory layers for AI agents, and they made opposite architectural bets — vector-first versus temporal knowledge graph. Both left a hole where open temporal memory should be. Here is the honest map, numbers included.
Zep vs Mem0 vs OpenZync: The Honest Map of Agent Memory in 2026
engineering - 2026.08.03Most agent-memory tools are vector stores with a retrieval loop. OpenZync makes a different bet: facts are versioned rows with validity windows, conflicting facts supersede each other transactionally, and graph edges stay consistent with the fact store. This post walks the full architecture.
Agent Memory as a Temporal Knowledge Graph — A Technical Deep Dive
engineering - 2026.08.02A support agent that forgets what a customer said three weeks ago isn't broken — it's just stateless. A practical look at how vector, graph, and temporal memory behave in production, through a real agent build.
AI Agent Memory in 2026: Why Context Windows Aren't Memory
engineering - 2026.07.16Today we're announcing OpenZync: production-grade, graph-based memory infrastructure for AI agents with support for 3 graph backends and 5 LLM providers.
Welcome to OpenZync — Persistent Memory for AI Agents
releases