Memory: how the platform learns

What kinds of memory does an agentic platform need?

• →Episodic memory: what happened — incidents, actions taken, outcomes, time-to-resolution. The raw material of learning.
• →Semantic memory: how things work here — architecture decisions, service ownership, environment constraints, "we never deploy X on Fridays."
• →Procedural memory: what to do — refined remediation sequences that worked, golden-path customizations, escalation paths.
• →Working memory: the current task's context window — assembled fresh from perception and retrieved long-term memory per request.

Why must memory be scoped?

Unscoped memory is a liability multiplier. A resolution that was correct for one tenant may be wrong for another; a remediation that worked at one scale can be destructive at another. Memory needs tenant, service, environment, and principal scoping — plus retention and deletion semantics, because "the platform remembers everything forever" is a compliance statement you did not mean to make. Remembering is a governed capability, sibling to identity, not a free-form cache.

How does memory interact with learning?

The loop is: act, observe outcome, record. When an agent's remediation succeeds, the sequence and its context are written to memory; the next similar incident retrieves it as preferred context. When it fails, the failure and its correction are recorded too — negative memory is often the more valuable half. Over time this compounds into what the platform "knows," which is precisely the knowledge as a product shift: documentation stops being a wiki humans maintain and becomes a byproduct of governed operation.

What are the failure modes?

• ✕Memory without curation. Recording everything produces retrieval noise; the agent drowns in its own history. Memory needs promotion and decay, like any knowledge base.
• ✕Stale certainties. Infrastructure changes; last year's "correct" fix becomes this year's outage. Memory entries carry freshness metadata and confidence, and old lessons get re-validated.
• ✕Cross-contamination. Unscoped memory lets one team's constraints leak into another's decisions. Scoping is not optional.