Introducing the Organisational Memory System: the missing layer in enterprise AI
AI agents are becoming more established and widespread in organisations, but the proliferation is moving faster than the systems to govern them. A missing piece of the puzzle is a durable, auditable memory of how an organisation wants its agents to behave, kept current and ready to inspect. Whilst this is available for individual users, teams and organisations are not well-served. Here is where the market stands, what is still absent, and what we are doing about it.
Most organisations now run AI agents somewhere: drafting replies, triaging tickets, writing code, assembling research. What has not kept pace is the ability of those agents to learn from the organisation they work for.
Correct an agent today and the correction holds for that conversation. Tomorrow a different agent, or the same one in a fresh session, repeats the mistake. "Use our approved travel provider, not the consumer site." "Never paste customer data into an external tool." Each correction is useful once, then lost when the session ends. The organisation pays to discover the same lesson again and again (in time and in tokens).
An example: the branded, accurate slide deck
Take a task common to many organisations: producing a slide deck that uses the right product names, the current positioning language and tone, the approved logos and brand assets, the agreed value propositions.
This information can be captured in a "Skill", which is a great baseline. However, the desired behaviour this skill dictates will inevitably morph over time, and multiple users are likely to challenge and tweak the outputs in their own way to get to a satisfactory result. Or, perhaps newer models perform differently and extra instructions are required to nudge things in the right direction.
There is no collective knowledge of these user tweaks and the evolution of requirements.
What is needed is a system where the agent producing the deck knows two things automatically: what "good" looks like for this organisation right now, and what this specific salesperson tends to get wrong. The first is shared institutional knowledge (the codified layer). The second is private behavioural memory (the recollection layer). Both feed the same task, both are kept current without anyone manually editing a knowledge base, and the agent that produces tomorrow's deck is incrementally better than the one that produced yesterday's.
This is a memory problem, though not the one most people picture when they hear "AI memory".