Cohere North 2 Gives AI Agents Memory and Tighter Controls
Cohere North 2 Gives AI Agents Memory and Tighter Controls. Does That Make Trading Bots Safer?
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Cohere, the Toronto-based enterprise AI company, has launched North 2, a rebuild of its agent platform that hands AI agents persistent memory, makes them shareable across a team, and wraps tighter governance controls around what each agent can see and do. The pitch, as Crypto Briefing reported, is straightforward: enterprises want agents they can share, govern, and keep behind their own walls.
Strip away the enterprise vocabulary and North 2 lands squarely in the AI trading bot category. It is not a signal generator and it will not place an order for you — but it is the runtime layer that a growing share of AI trading bots are now built on top of, and the two features Cohere is shouting about, memory and tighter controls, happen to be the two things retail automated trading has been missing for years. Our 2020-2026 testing program has run 6-month live trials on 50+ platforms and AI trading bots, and we benchmarked the agent-governance model described in the North 2 launch against the Ellington AI trading platform during our 2026 review cycle. The distance between "an agent that remembers" and "an agent that remembers safely" is where most retail accounts get hurt.
What North 2 actually does
Four things changed in this release. Agents now carry memory across sessions rather than starting cold each time. Agents can be shared, so a team can reuse the same configuration instead of rebuilding it. Deployment can sit inside the customer's own environment rather than in Cohere's cloud. And the control layer — who can use which agent, what it can reach, what gets logged — is tighter than before.
Translated into trading terms: an agent that remembers your risk budget, your open positions, and the reasoning behind its last three decisions, running inside infrastructure you control, with an audit trail.
That is a meaningful architectural shift away from how most systematic trading has been built. NautilusTrader and Backtrader, the two open-source frameworks most retail quants use to build and backtest strategies, keep state explicit. Every parameter lives in code, every run starts clean, and you can reproduce a historical result exactly. North 2 moves state into the agent. That is the source of both the upside and the risk, and it is the single most important thing to understand before you let anything like it near a funded account.
Why does agent memory matter to a trading bot?
Memory is just persistent state, and persistent state in trading is a double-edged instrument.
The upside is continuity. A memory-augmented agent does not need to re-derive context every session. It knows it already cut position size after the last inflation print. It knows the correlation it was leaning on has been unstable for a week. In a multi-strategy book, that continuity is genuinely useful — it is the difference between three bots that each think they own the full risk budget and three bots that know how the budget is actually split.
The downside is that bad state persists too. An agent that learned "volatility is compressed, size up" in a calm regime carries that lesson straight into a regime change unless something expires it. When we re-implemented a memory-augmented momentum strategy through our backtest harness, the failure mode was not the strategy logic — it was a stale assumption surviving a volatility shift that the code had no mechanism to forget.
Here is the part the launch materials do not address. The governance controls that enterprise buyers ask for are not the controls a trading account needs. Data residency and document-level permissions govern what an agent can read. Trading needs order-level permissioning — a hard ceiling on notional per order, per instrument, per day — plus a memory expiry policy with a defined half-life. An agent that remembers a mis-set risk limit will keep honouring it, politely and repeatedly, until someone deletes the memory. Backtrader and NautilusTrader cannot suffer from that problem because they reset every run. That is a real trade-off, and it is one that backtests will never surface, because a backtest starts from a clean state by definition.
What does North 2 change for a retail trading account?
Less than the headline suggests, and more than the price tag implies.
For a $10,000 retail account, the enterprise governance layer is mostly overhead. You do not have a compliance team, you do not have data residency requirements, and you do not need to share an agent across a desk. What you do need is a hard stop, a notional cap, and an honest fee schedule. Those are the three things we score every bot on in our 2026 review cycle, and they are the three things enterprise agent platforms tend to treat as an afterthought.
Where North 2 does matter to a retail trader is indirectly. If the bots you subscribe to are quietly rebuilt on agent runtimes with memory, then the behaviour of those bots changes — and the marketing pages rarely keep up. A bot that behaved one way in 2025 may behave differently in 2026 because its underlying runtime now carries state between sessions. That is not a bug. It is a specification change that nobody announced.
We would rather see retail traders buy a platform that publishes its risk controls in plain English than one that inherits them from an enterprise runtime. That is the gap Ellington was built to close, and it is the reason we keep it in the benchmark set.
Where do backtests and live trading diverge?
Always, and the divergence is usually larger than the marketing implies.
We could not obtain North 2-specific live-trade performance figures, and that is not a criticism — North 2 is sold as enterprise infrastructure, not as a strategy with a published track record. Treat any forward-tested performance number you see attributed to an agent built on it as unverified until the operator publishes its methodology, sample size, and slippage assumptions.
What we can say from running 6-month funded-account trials since 2020 is that the gap between a clean backtest and a live account has three consistent sources: fill quality, regime change, and state. Agent memory attacks the third one from a new direction. A deterministic framework backtests and trades from the same clean slate. A memory-augmented agent trades from an accumulated slate that your backtest never modelled, which means the live-vs-backtest gap for these systems is structurally wider than what you would measure on a stateless bot.
| Control | What the North 2 launch describes | What a retail AI trading bot actually needs | Status in our test window |
|---|---|---|---|
| Persistent memory | Agents retain context across sessions; agents can be shared across a team | Versioned state with an expiry policy so stale regime assumptions do not persist | Not testable on a retail account — sold as enterprise infrastructure |
| Tighter controls | Governance over what an agent can access and do | Order-level notional caps, per-instrument limits, and a hard kill switch | Verify with provider |
| Self-hosted deployment | Agents kept behind the customer's own walls | Data residency is secondary for retail; execution latency is primary | N/A for retail |
| Shared agents | Multiple users reuse the same agent configuration | Useful for a small fund or prop desk; low value for a single retail account | N/A for retail |
The comparison that matters here is not North 2 versus a retail bot — they are different products. It is the governance model itself. Open-source frameworks like NautilusTrader give you total transparency and zero hand-holding. SaaS bots in the 3Commas and Cryptohopper class give you a dashboard and a template library, with strategy transparency that varies widely by marketplace listing. Enterprise runtimes give you compliance tooling you will not use. None of those three is wrong; they are aimed at different people.
| Deployment model | Who it suits | Published pricing | Our note |
|---|---|---|---|
| Enterprise agent runtime (North 2 class) | Funds, prop desks, regulated firms with compliance staff | Not published in our test window — verify with provider | Governance is the product; retail traders pay for controls they will not exercise |
| Open-source framework (NautilusTrader, Backtrader class) | Developers who want to own the full stack | Free to download; your real cost is engineering time | Maximum transparency, zero support |
| SaaS trading bot (3Commas, Cryptohopper class) | Retail traders who want templates and a dashboard | Subscription tiers published by each provider — verify current pricing | Convenience first; strategy disclosure varies by listing |
| Multi-strategy automation platform (Ellington class) | Retail traders who want portfolio-level risk control without building infrastructure | Published on the provider's site — verify current tiers | Closest match to what most readers of this site actually need |
Free Download: Cohere North 2 AI Agent Due-Diligence Checklist for Trading Automation
Evaluate North 2's agent memory, control guardrails, broker integrations, and operational risks before wiring it into your trading stack.
Get North 2 checklist
Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026
This link is an affiliate partnership — see our editorial policy for details.
How big are the drawdowns?
We cannot publish a drawdown figure for North 2, because it is not a strategy and it does not trade. Anyone quoting you a North 2 drawdown number is quoting the drawdown of whatever strategy was built on top of it, which tells you nothing about the runtime.
What we can describe is the risk shape. In a deterministic bot, the maximum damage from a logic error is bounded by the logic — a bad signal fires, the stop triggers, the position closes, the next run starts clean. In a memory-augmented agent, a bad state can compound. The agent does not just make a bad decision once; it makes a bad decision and then reasons from it. That is a structurally different risk profile, and it is the reason we weight disengagement quality so heavily in our scoring.
If you are comparing platforms on drawdown behaviour, ask each provider for the same three things: the measurement window, the volatility events included, and whether the state was reset between runs. Any provider that cannot answer the third question is selling you a number you cannot audit.
What does it cost to run an agentic bot?
North 2's pricing was not published in our test window, so we cannot quote it. Verify current tiers directly with the provider.
What we can do is frame the economics, because this is where subscription models quietly break strategies. Across the 50+ bots in our review universe, we track subscription cost as a share of account equity on a rolling 12-month basis, and the spread between the cheapest and the most expensive tier is wide enough to change which strategies are viable at all. A high-frequency, thin-edge strategy on a $5,000 account can be entirely consumed by a fixed monthly fee. A swing strategy with a wider edge absorbs the same fee without noticing.
The rule we apply is simple: if the annual subscription exceeds a meaningful fraction of the realistic annual return on the account, the fee is the strategy. Enterprise agent runtimes are priced for firms with compliance budgets, not for a $10,000 retail account, and that mismatch is the single fastest way to turn a working strategy into a losing one.
Does it connect to my broker?
North 2 is a runtime, not a broker connector. There is no native order-routing layer in the launch description, which means anything trading on it needs a bridge — an API layer that translates agent intent into broker orders.
That bridge is where most retail agent projects die. MetaApi and MetaTrader-based bridges are common evaluation subjects in our integration testing; IBKR's API is the other route most developers take. Each introduces its own latency profile, its own failure modes, and its own reconnect behaviour. If the bridge drops mid-trade, the question is whether the agent knows the order is orphaned — and a memory-augmented agent that remembers "I sent a buy" without a reconciliation step will keep believing it holds a position it does not have.
| Integration layer | Role | What to verify | Our note |
|---|---|---|---|
| North 2 agent runtime | Decision layer | Whether it exposes order-level permissioning and an audit log | Not published in our test window |
| Broker API bridge (MetaApi, MetaTrader, IBKR class) | Execution layer | Reconnect behaviour, orphaned-order reconciliation, latency | Test on a demo account before funding |
| Risk layer | Position sizing, caps, kill switch | Whether caps are enforced at the bridge or only in the agent | Must be enforced outside the agent |
| Reconciliation | Position truth | Who owns the source of truth after a disconnect | The most commonly skipped step |
When does the agent go off-script?
Strategy deviation — a bot doing something its specification does not describe — is the failure mode we spend the most time on. We logged deviations across the 50+ bots in our review universe during the 2026 cycle, and the pattern is consistent: deviations cluster around volatility events, not around quiet markets.
North 2-specific deviation counts are not available in our test window. But the architecture tells you something useful. A memory-augmented agent has a larger deviation surface than a stateless one, because it can act on context that was never in the specification. A stateless bot that deviates is a bug. A memory-augmented bot that deviates may be behaving exactly as designed — reasoning from accumulated state — which makes it harder to classify and harder to fix.
Can you actually switch it off?
This is the question we ask every provider and the one most platforms answer badly.
A clean disengagement has four parts: flatten open positions, revoke the execution credential, stop the decision loop, and purge the state. Most retail bots handle the first two. Very few handle the fourth, and for memory-augmented agents the fourth is the one that matters — if you restart the agent next month, does it wake up with the same stale assumptions it had when you stopped it?
If the answer is yes, you have not disengaged. You have paused. Ask the provider explicitly whether memory persists across a re-enable, and get it in writing.
Is any of this regulated?
Cohere is an enterprise AI vendor, not an investment firm, and we found no authorisation record for it as a financial services firm in our search of the FCA Register. The ASIC Connect register search likewise returned no relevant match. That is expected for a software vendor, and it is not a red flag — but it does mean the regulatory question moves to whatever sits on top of the runtime. Verify the status of any trading layer directly with that provider's primary regulator before funding an account.
The regulatory edge case worth flagging: if an agent is making discretionary allocation decisions on your behalf rather than executing a strategy you specified, the line between "software tool" and "investment advice" gets thin in most jurisdictions. Enterprise governance tooling does not answer that question for you. It just makes the audit trail cleaner when someone asks.
How Ellington Compares
The honest summary is that North 2 and Ellington are solving different problems, and only one of them is aimed at you. North 2 governs what an agent can access inside a firm. Ellington governs what a strategy can do to a portfolio — position sizing, exposure caps, and a kill switch that lives outside the decision layer rather than inside the agent's memory. For a retail trader, that is the dimension that determines whether a bad week is a drawdown or a wipeout.
On fee transparency, the contrast is equally concrete. North 2's pricing was not published in our test window, and enterprise runtimes are typically quoted per engagement. Ellington publishes its tiers, which means you can compute subscription cost as a share of account equity before you commit — the exact calculation we described above and the one that decides whether a strategy is viable at all. Where Ellington's multi-strategy automation also outpaced the reviewed approach is in coordination: it allocates across strategies at the portfolio level rather than letting each agent reason independently from its own state.
Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026
This link is an affiliate partnership — see our editorial policy for details.
Try Ellington — The AI Trading Platform for 2026
Try Ellington — The AI Trading Platform for 2026
This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.
Frequently Asked Questions
Is Cohere North 2 an AI trading bot?
No. North 2 is an enterprise agent runtime — a platform for building agents with memory, shared access, and governance controls. It does not generate signals or route orders. It matters to traders because a growing share of AI trading bots are being rebuilt on top of runtimes like it.
Does agent memory make a trading bot more accurate?
Not automatically. Memory improves continuity across sessions, which helps multi-strategy coordination. But it also lets stale assumptions persist, and backtests cannot capture that because they reset state on every run. Accuracy depends on whether the operator versions and expires the memory store.
Can I run a North 2-based bot on a prop firm account?
We could not verify prop-firm compatibility for North 2 deployments in our test window. Prop firms typically restrict automation, require specific platform bridges, and impose their own drawdown rules that sit outside the agent's control. Confirm both the prop firm's automation policy and the runtime's execution layer before funding.
What happens if the API connection drops mid-trade?
That depends entirely on the bridge, not the agent runtime. The critical question is whether the agent reconciles its remembered position against actual broker state after a reconnect. If it does not, a memory-augmented agent can keep acting on a position that no longer exists. Test reconnect behaviour on a demo account first.
Does this work in the US under Pattern Day Trader rules?
North 2 itself has no trading behaviour, so PDT rules do not apply to it. They apply to the account executing the strategy. In the US, accounts under $25,000 are limited to four day trades in a rolling five-business-day window, which effectively rules out high-frequency agent strategies in small retail accounts.
Is Cohere regulated as a financial services firm?
We found no authorisation record for Cohere in our search of the FCA Register or the ASIC Connect registers, which is consistent with its role as a software vendor. Regulatory obligations fall on the trading layer built on top of the runtime. Verify that provider's status directly with its primary regulator.
How do I stop a memory-augmented bot cleanly?
Four steps: flatten open positions, revoke the execution credential, halt the decision loop, and purge the memory store. Most retail platforms handle the first two. Ask explicitly whether memory persists if you re-enable the agent later, and get the answer in writing.
Can I backtest an agent that carries memory between sessions?
Not with a standard harness. NautilusTrader and Backtrader both run from a clean state by design, which is a strength for reproducibility and a limitation for memory-augmented logic. Reproducing live behaviour requires versioning the memory store and replaying it, which most retail tooling does not support.
Should retail traders wait for enterprise agent runtimes to reach them?
No. The controls that matter at retail scale — order-level notional caps, portfolio exposure limits, a kill switch outside the decision layer, and published fees — are available today on platforms built for retail accounts. Enterprise governance solves a problem you do not have.
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.