Every morning, our AI assistant Jarvis wakes up with no memory of yesterday. No recollection of the emails it processed, the decisions it helped make, or the context it spent hours building up. It's like hiring a brilliant employee who develops amnesia every night.
This is the reality for every business running AI agents today. And it's a bigger problem than most people realise.
The Amnesia Problem
At Drakon Systems, we use AI agents internally to handle real operations — monitoring systems, processing emails, managing infrastructure, and coordinating workflows. Our assistant Jarvis runs on Claude and handles dozens of tasks across the business day.
The problem became obvious within the first week. Jarvis would have a detailed conversation about a client project in the morning, build up rich context about priorities and preferences, then lose all of it by the afternoon session. Every interaction started from zero.
The typical workaround is to stuff the AI's context window with everything it might need — conversation logs, documents, notes. But context windows have hard limits. You can't feed an AI six months of operational history every time you ask it a question. The costs alone would be astronomical, and the signal gets buried in noise.
What we needed wasn't a bigger clipboard. We needed actual memory.
How Human Memory Works (And Why It Matters)
Human memory doesn't work like a database. You don't store every experience with equal weight and retrieve them by running queries. Instead, your brain does something far more sophisticated:
- Salience detection — emotionally significant or unusual events get flagged for stronger encoding
- Temporal decay — memories naturally fade unless reinforced through recall or relevance
- Consolidation — during sleep, your brain reorganises short-term memories into long-term storage, strengthening important connections and pruning the rest
This is why you remember a critical client meeting from three months ago but not what you had for lunch last Tuesday. Your brain is constantly curating, prioritising, and reorganising.
We realised that if we wanted Jarvis to have useful memory, we needed to model these same principles — not just log everything, but intelligently decide what matters and surface it at the right moment.
Enter Claude Cortex
Claude Cortex is the memory system we built to solve this. It's named after the cerebral cortex — the part of the brain responsible for memory, attention, and thought — because it's designed to work the same way.
Here's how it works:
Salience Detection
When Jarvis processes information — a conversation, an email, a task outcome — Cortex evaluates how important that information is. It considers factors like emotional weight, novelty, and relevance to ongoing work. A routine status update scores low. A client escalation or a key decision scores high.
High-salience memories get encoded more strongly and persist longer. Low-salience ones fade naturally, just as they would in a human brain.
Temporal Decay
Every memory has a strength score that decays over time. Recent memories are vivid and easily accessible. Older ones fade unless they're reinforced — by being recalled, referenced, or connected to new information. This means Cortex naturally keeps fresh, relevant context at the top without manual curation.
Memory Consolidation
Periodically, Cortex runs a consolidation process analogous to what happens during sleep. It reviews accumulated memories, identifies patterns and themes, merges related memories into stronger consolidated representations, and prunes noise. After consolidation, the memory store is cleaner, more organised, and more useful.
Contextual Retrieval
When Jarvis needs to recall something, Cortex doesn't just do keyword search. It combines semantic similarity, recency, and importance to surface the most relevant memories for the current context. Ask about a client, and Cortex recalls not just the last interaction but the pattern of interactions — preferences, issues, decisions — weighted by relevance and recency.
Privacy-First and Open Source
We made two deliberate architectural decisions that we think matter for any business considering AI memory:
Local-only storage. Cortex uses SQLite — everything lives on your own infrastructure. No memories are sent to external services, no cloud databases, no third-party access. For a system that's literally recording what your AI learns about your business operations, this isn't optional. It's essential.
Open source. We released Claude Cortex as an open-source project because we believe memory is foundational infrastructure for AI agents. It shouldn't be locked behind proprietary platforms. Any team building with Claude — or any LLM — can use, audit, and extend it.
The Missing Piece for Business AI
Most conversations about AI in business focus on capabilities: what can the model do? Can it write emails, analyse data, generate reports? These are important questions. But they miss a fundamental issue.
Without memory, every AI interaction is isolated. Your AI assistant can't learn your preferences. It can't build on previous work. It can't notice patterns across weeks or months. It can't develop the kind of institutional knowledge that makes a human team member increasingly valuable over time.
Memory is what turns an AI tool into an AI teammate.
Since deploying Cortex, Jarvis maintains context across sessions, remembers decisions and their rationale, recalls client preferences without being reminded, and builds on previous work rather than starting from scratch. The productivity difference isn't incremental — it's transformational.
What This Means for Your Business
If you're building with AI agents — or planning to — consider memory as a first-class requirement, not an afterthought. The models will keep getting smarter. But without persistent, intelligent memory, they'll keep forgetting.
Claude Cortex is available on GitHub for teams who want to add memory to their own AI workflows. It's free, it's private, and it works.
At Drakon Systems, we build AI-powered tools that solve real operational problems. Our AI invoice processing platform uses the same philosophy — intelligent automation that learns and adapts, not rigid templates that break. If you're looking to bring AI into your business workflows, get in touch.
The future of AI isn't just smarter models. It's models that remember.