Updates, insights, and deep dives from the Caura team.
Looking for case studies and integrations? Browse Use Cases →
Your company’s knowledge changes over time. Your agents should understand that. As-Of Recall aligns memory search with the date a question refers to — so “What was our pricing policy last March?” brings forward what was true in March. By measuring freshness from recorded event dates instead of upload dates, imported history gets a real timeline and agents answer historical questions with the right context. One tenant setting, no new fields, off by default.
Read article →At enterprise scale, orchestration is a memory problem before it is a routing problem. The four control patterns and where each degrades past a dozen agents, an audit of what seven frameworks actually document about cross-agent memory (none mentions an audit trail), the four governed-memory mechanisms that fix the failures, and four runnable Apache-2.0 reference implementations — plus what the whole thing costs to operate.
More agents buy you parallelism and collisions at the same time. The four ways fleets step on each other — duplicate work, constraint contradictions, stale facts, boundary leaks — where orchestrator-worker, handoffs, A2A/MCP messaging and plain vector stores each give out, and the five mechanisms that make a shared store actually enforce coordination. Plus an original finding: in the reference fleet, coordination is enforced 14 times, and never by a prompt.
Persistent memory only works when four things hold at once: the fact can be found, is current, traces back to its writer, and is visible only to agents allowed to see it. Most implementations ship the first and skip the rest. The write and recall mechanics, a definition audit of the three founding papers (zero governance primitives in 67 pages), five production failure modes including memory poisoning, and three repos that reproduce each one.
AI agent memory is a store outside the model with a write path, a read path, and a lifecycle — the context window is a working set, and RAG is a read path over documents. What persistence costs, the four failure modes it creates (scope, time, resolution, provenance), what governed recall runs before it ranks, and the measured production numbers — including the two failures Caura published about its own service.
Diligent journaling competes with the task for the same attention, and it loses. The Interviewer is Caura's reflective write path: on a schedule it reads an agent's existing work trail — a Claude Code or Cursor transcript, or an OpenClaw plugin buffer — and synthesizes typed, governed memories after the fact. Two write paths, one governed layer; coverage without the journaling tax. The six-section mapping, the plugin-buffer and disk-parser collectors, the watermark protocol, and why it ships off by default and double-gated.
Memory as a model behavior is a reliability bug: the model forgets to write, writes twice, or writes what a poisoned document told it to — and none of it throws. Deterministic memory moves every write and the baseline read into the harness, where policy is a permission system instead of a prompt. What determinism guarantees (invocation and governance, not correctness), the failure modes it fixes vs. relocates, what the OpenAI Agents SDK, LangGraph, ADK, Zep, Mem0 and Letta actually shipped — verified against their docs — and a 130-line demo where a Gemini agent writes and a Claude agent recalls from one governed store.
There are only five ways agents share knowledge: message passing, orchestrator state, shared files, a shared vector store, or a governed memory layer. The first four move information; only the last compounds it. The five architectures and where each breaks, the four-step loop that turns one agent's failure into fleet policy, and the wiring for Claude Code, CrewAI, AutoGen, and LangGraph.
Single-agent memory is a nicer chat history. A fleet needs something else: a shared, governed store where every item carries scope, provenance, trust and validity. The taxonomy, the six write gates, why over-retrieval hurts more than forgetting, the five failure modes — and the cases where you should skip all of it.
Five agents, twelve live data sources, one governed memory. Caura's growth function stopped being people assembling dashboards and became a fleet that remembers — Beacon on analysis, Outreach on the funnel, Social on engagement, Scout on the outside-in radar, Writer on content. The three properties that separate a department from a demo: one tool surface, governed shared memory, and a human gate automation never widens. Plus the four things still broken in our own store.
Block open-sourced Buzz, an Apache-2.0 workspace where every participant — human or agent — holds their own Nostr keypair instead of an API key managed by a vendor, every action lands as a signed event under a hash-chain audit log, and every agent carries its own encrypted engram (NIP-AE). Buzz ships more memory than most agent platforms. An appreciation of what it gets right, plus our initial research into the third kind a fleet needs: the shared, governed tier beside the private one.
PeerRank's blind run “July25” put five frontier models across 100 questions and 2,922 pairwise matches. Claude Opus 5 won outright at 8.87, leading four of five categories. Claude Fable 5 finished third — four answers came back blank, HTTP 200 with an empty body, and the judges scored what they saw. Plus kimi-k3: second on quality, 18.81 seconds per answer, and a judge panel whose disagreement about how to mark was four times larger than the gaps it was marking.
In PeerRank's blind run “Mondial,” Claude Fable 5 posted the highest head-to-head win rate of four frontier models — then finished third, because its safety layer refused four ninth-grade biology questions and logged the blanks as empty, successful calls averaged into its score. The numbers, the forfeits, and why refusal behavior belongs in fleet selection criteria.
A brand-new agent has flawless reasoning and nowhere to stand. Pre-seeded, scoped ingestion (per organization, per department) plus mandatory keystones give it the knowledge base and the rulebook on turn one — governed, auditable, and shared, instead of an ever-growing system prompt.
Claude Fable 5 posts the highest win rate on PeerRank's board — then places third, because a safety classifier refuses ninth-grade biology and logs the refusals as empty, successful calls that get averaged into its score. The numbers, the forfeits, and the fix Anthropic already ships.
Our new arXiv paper formalizes the fleet-memory problem, defines the primitives a governed memory system needs, and measures Caura against a live production service — including the two architectural bugs the measurement caught. The negative results are the point.
When several agents independently learn the same lesson, Caura's Skill Factory distills it into a reusable skill — then a deterministic scanner and an active-only gate keep it safe. The mechanism, plus a live run that blocks 6/6 adversarial skills.
Most teams build organizational intelligence as a pile of bespoke skills — one per capability, one per agent. You don’t need the pile. You need one skill, used properly, over governed shared memory: recall before work, obey the keystones, reuse the playbooks, compound what every agent learns.
In a fleet, the tokens that dominate the bill aren’t spent on reasoning — they’re spent on repetition. The memory-infrastructure principles that keep cost flat as the fleet grows.
When your user pushes back and your AI agent caves, the problem isn’t the model — it’s the enforcement layer. Probabilistic enforcement isn’t enforcement; it’s hope. Here’s how Caura’s keystones primitive fixes it.
Apache 2.0. The whole storage layer, the 12 MCP tools, the OpenClaw plugin, the audit trail — yours to read, run, fork, and ship. Five minutes from git clone to a working multi-agent memory layer.
Six operations and one collection-based primitive that replaces a shelf of side-systems. Customer records, config, skills, playbooks — one tool, with semantic search opt-in per collection.
Single-agent memory is a solved category with many good vendors. Multi-agent governed shared memory is a new category — and Caura is the one defining it.
Caura on the two public agent-memory benchmarks: 23 ms p50 search, 96–99% token savings, accuracy comparable to the leaders — and the fleet-shaped problem these benchmarks can’t measure.
The Karpathy Loop proved autonomous AI research works. But scaling it to agent fleets needs governed shared memory — persistent, structured, and self-improving.
How AI agents evolved from stateless chatbots to Karpathy loops and Meta’s self-modifying hyperagents — and why governed shared memory is the missing infrastructure layer.
How the governed agent DB for AI fleets is built: three memory modes (Rail, MCP/REST, the Interviewer) over one governed write path and recall path, the Evolve loop that makes recall improve from outcomes, keystones and trust tiers, and the stack underneath. Updated September 2026.
OpenClaw turned AI from a tool you prompt into a coworker that lives on your machine. Now enterprises are deploying fleets — and discovering that the hardest problem isn’t the agent.
Agent fleets are scaling. Memory isn’t. The missing layer between isolated agents and compounding intelligence is governed shared memory — and building it is harder than you think.