For teams on Claude Code, Codex and Cursor
Give Claude Code the right memory on every turn.
More accurate answers, fewer hallucinations and a smaller AI bill. elephant.broker decides what your agent sees on every turn, across Claude Code, Codex, Cursor and your own harness.
+36%
better accuracy
stale-state errors fell from 47% to 11%
+21%
more tasks completed
71.0% to 91.6% on the same 50 tasks
60%
faster task runtime
60.2% less wall-clock time, same workflow
- PlacedRefund policy v4 (current), the step this task is on
- DroppedRefund policy v3 (superseded)
- Dropped14 tool results this task no longer needs
- FlaggedCLAUDE.md and AGENTS.md disagree on deploy steps · weighed by proof
Your model reads the working set, not the whole history.
The problem
Your agents don't share context or memory.
No memory between sessions
Every session starts from zero. The decision, the correction, the procedure that worked yesterday: gone by the next one.
No context shared across agents
Your tools, teammates and other agents each hold a piece of what your company knows. No agent sees the rest, so each one relearns it or guesses.
No shared record of what is current
With nothing shared to check against, two conflicting instructions are a coin toss: Claude “may pick one arbitrarily”. Your agents guess, and can't tell you why.
So every agent fills the gap the same way: it re-sends the whole history, every rules file and every tool result on every turn. You pay for all of it, and a bigger window doesn't fix it.
How it works
We don't host your agents. We change what they see, every turn.
- 1
Assemble
Pulls task-relevant context from your tools, teammates and past sessions.
- 2
Prune
Keeps what is current and verified. Drops stale, duplicate and off-task context.
- 3
Place
Puts the working set and the current procedure where the model pays attention. Conflicts are flagged and weighed by proof.
- 4
Write back
Captures decisions and corrections, so the next turn starts from them.
Your model and your harness stay unchanged. Connect through an MCP server, native plugins or a REST API: Claude Code, Codex, Cursor, OpenClaw, Hermes or your own harness.
Built on our own research: ElephantBroker: A Knowledge-Grounded Cognitive Runtime for Trustworthy AI Agents (arXiv 2603.25097).
Results
What changes when your agents read less
+36%
better accuracy
stale-state errors fell from 47% to 11%
+21%
more tasks completed
71.0% to 91.6% on the same 50 tasks
40%
lower AI spend
same model; savings compound on long-running work
40-80%
less context per task
only verified, relevant knowledge reaches the window
60%
faster task runtime
60.2% less wall-clock time, same workflow
71%
fewer tasks deferred
71% fewer non-completions, with 62.7% fewer tokens
Measured per task on the same model and workflow.
How it compares
The car, not another engine.
Memory providers are the engine. elephant.broker is the car that uses one to get you from A to B. Cognee is the default memory. Already running another provider? Plug it in and keep it.
Scroll the table sideways to see every column.
| Versus | They | elephant.broker |
|---|---|---|
| Memory providers | Store and retrieve facts. The engine on its own. | Uses your memory to decide what enters the window on every turn. |
| Frontier lab memory | Works only inside that vendor's own tools. | One context across every harness and model your teams use. |
| Context warehouses | Store company context, like a data warehouse. | A runtime: decides what each turn sees, from pointers. |
| Connectors and enterprise search | Connect to Slack, email and Notion and see everything. | Uses pointers, not the data, and picks only what each turn needs. |
| AI cost tools | Switch to a cheaper model or compress the transcript. | Same model, smaller working set, more accurate answers. |
Book a call
Ready for more accurate agents and a smaller AI bill?
Security
Nothing moves.
No copies, no connectors, no keys
We keep pointers. Your agents use the access you already gave them.
Single-tenant by default
EU regions, your own VPC, on-prem or air-gapped.
Zero data retention
Your data never trains anyone's model.
A record of why
A signed, append-only log of every read, placement and write-back.
The team
The team behind elephant.broker
Sarunas Legeckas
CEO & Co-founder
- Operator
- Ex-N26 GM across 18 EU markets.
- Go-to-market
- Runs enterprise pilots, partnerships & rollout.
Alexandru Lupascu
CTO & Co-founder
- Engineering
- Ex-CTO & co-founder @ OMNIA Protocol. Patent EP3883204B1.
- AI runtime
- Ships the agent runtime, integrations & audit engine.
Cristian Lupascu, PhD
CPO & Co-founder
- AI research
- Ex-CEO & co-founder @ OMNIA Protocol. PhD in privacy-enhancing technologies.
- Product & Security
- Designs the product features & security.
Simas Simanauskas
Founding partner, GTM
- Commercial
- Ex-Chief Commercial Officer @ ConnectPay. Drove 18x client growth.
- Go-to-market
- Leads sales, design partners & key accounts.
Before you book
Questions before a call
Will security approve a pre-seed vendor?
We hold no copies, no connectors and no keys. Single-tenant, EU, VPC, on-prem or air-gapped. A one-page brief goes to your reviewer before any data is touched. SOC 2 and ISO 27001 are planned, not held.
Does it add latency?
Yes: a retrieve and rerank step on every turn. The working set is smaller, so the model reads less. You see both numbers on your dashboard.
Won't Anthropic or OpenAI build this?
Their memory works only inside their own tools. Most teams run Claude, ChatGPT and Gemini side by side, and each vendor's memory stays inside that vendor. elephant.broker gives all of them one context.
We could build it ourselves.
A hook that injects the right rules is the right idea. The work is everything around it, across every harness, while the tools change monthly. Run both on your own tasks and compare. If yours wins, keep it.
We already run a memory library or enterprise search.
Keep it. elephant.broker comes with Cognee as the default, and any other memory provider plugs in easily, with all the benefits: per-turn decisions, more accurate agents and lower AI spend. Search finds things. We decide which of them belong in this turn.
Will we really save 40%?
Measure it. Run your own tasks with and without elephant.broker on the same model and compare the bill.
What if you disappear?
Everything exports: nodes, edges and traces, and it runs self-hosted. Leaving us is a file, not a migration.