Blog
Field notes on coordination memory
Category definitions, honest benchmarks, and build logs — written in the open.
Agent memory: episodic, semantic, and coordination
“Agent memory” is one phrase doing three jobs. The split cognitive science drew — episodic and semantic — plus the type multi-agent systems actually need: coordination memory.
Read the article →Per-agent access control for AI agents, compiled to SQL
When agents share memory, “who can read what” becomes a security boundary. Prompt-level rules are advisory; a policy compiled to a SQL predicate is enforced in the query plan.
Read the article →Deterministic context packs and how prompt caching compounds
Multi-agent token bills balloon because every agent re-sends history. The underused lever is determinism: identical inputs → identical bytes → prompt-cache hits that compound across the fleet.
Read the article →Claude Agent SDK: giving your agents shared memory
The Claude Agent SDK builds one capable agent fast — but its agents don’t share what they learn across runs and subagents. Point them at Lore’s MCP server and the whole fleet works from one reality.
Read the article →Context engineering for multi-agent systems
Context engineering is the craft of deciding what goes in the window. For one agent it’s curation; for a team of agents it becomes a distributed-systems problem — and the tactics that work solo quietly break.
Read the article →AI agents forget context — the fix isn’t a bigger model
It looks like forgetting, so teams reach for a bigger context window or a smarter model. But a multi-agent system that loses context has a coordination problem, not a capacity one — and scale doesn’t fix it.
Read the article →Why multi-agent systems fail — reading the MAST taxonomy
A 1,600-trace study found most multi-agent failures aren’t reasoning errors — they’re coordination errors. The taxonomy, and what it says your stack is missing.
Read the article →Multi-agent memory: why your agents live in parallel universes
Everyone says “agent memory,” but they don’t mean the same thing. A field guide to coordination memory — the category your multi-agent system is missing.
Read the article →Read-your-writes for agents, explained
The oldest trick in distributed databases, applied to agent handoffs — the one consistency guarantee that makes them stop failing by luck.
Read the article →The agent-memory landscape in 2026: Mem0, Zep, and the gap
Three products say “agent memory” and solve three different problems. An honest map of Mem0, Zep, and the coordination axis they leave open.
Read the article →AI agent memory benchmarks in 2026: the same-judge rule
One vendor reports 94.8; an independent run measures the same tool at 49.0. Why the numbers are nearly meaningless out of context — and the rule that fixes them.
Read the article →Memory poisoning in AI agents — and how quarantine stops it
Shared memory is a shared attack surface. How trust tiers, quarantine, and access control keep a poisoned write out of every agent’s prompt — by construction.
Read the article →What is coordination memory? A working definition
“Agent memory” is one phrase doing three jobs. A precise definition of the third one — coordination memory — and how it differs from just sharing a store.
Read the article →Agent memory on Postgres: how we built the write path
A write path that never blocks on an LLM, guarantees read-your-writes with a sequence number, and resolves conflicts with a database constraint — all on Postgres.
Read the article →No articles match — try another category or clear the search.
In the pipeline
- What is agent memory? A 2026 field guide
- Multi-agent state management without a message bus
- Agent handoffs that don’t lose context