2.1
Context resets
A new chat, model run, or coding session often starts without the source material that made the last answer useful.
SYMPTOM: repeated setup
Fig. 2 failure modes / category definition
FIG. 2
Why agents forget
Forgetting is usually not a model defect. It is a product boundary: the agent lacks a durable, scoped, inspectable layer for facts that should survive one session.
2.2
Sources drift
Docs, tickets, and decisions change. A memory layer needs supersession signals so stale facts do not look equally current.
SYMPTOM: stale recall
2.3
Evidence disappears
Raw retrieval can return snippets without enough lineage for a human to inspect what the agent relied on.
SYMPTOM: unverifiable answers
Fig. 3 definitions / not a sales page
FIG. 3
Memory vs RAG vs context
The terms overlap in conversation, but buyers evaluate different primitives. Memory is the persistence and governance layer around recall.
Temporary working memory inside one model call or chat. Useful, but it disappears or grows expensive quickly.
Search over source chunks. Useful for recall, but not automatically a durable, inspectable memory system.
A persistent layer that stores scoped evidence, provenance, updates, and deletion semantics for agents to reuse.
Fig. 4 trust requirements
FIG. 4
Why inspectability and supersession matter
Memory has to be inspectable because the user needs to know what was stored, why it was recalled, what source it came from, and whether a newer fact replaced it.
INSPECTABLE
Agent answers should expose the remembered source path instead of forcing humans to trust an opaque memory block.
SUPERSESSION
Changed facts need visible replacement signals so stale memory does not compete with current source evidence.
ERASURE
A memory product needs deletion semantics from the start. Stronger crypto-shred proof remains roadmap-labeled.
Crypto-shred proof Fig. 5 FAQPage schema / category answers
FIG. 5
FAQ
These category answers are short enough for AI answer engines to cite without turning the page into a sales deck.
What is agent memory?
Agent memory is a durable layer for grounded facts, source lineage, and recall across sessions and clients. It is different from one prompt or one search call.
Is memory the same as RAG?
No. RAG is a retrieval pattern. Agent memory adds persistence, provenance, supersession, and erasure semantics around what gets remembered and recalled.
Why does inspectability matter?
Inspectability lets humans see what was remembered, where it came from, whether it changed, and why an answer used it.
FIG. 6 - NEXT
Evaluate memory by what it lets you inspect.
Start with the product overview, then compare memory systems by inspectability, supersession, deletion posture, and tenant scope.
Memory: durable recall layer
RAG: retrieval pattern
Context: temporary working set
Trust: source lineage + supersession