Your AI Does Not Have Memory. You Are the Memory.

Most AI assistants do not have memory. They have history.
History is everything that happened. Memory is knowing what still matters.
The real workflow
Long-running AI work usually becomes this:
- paste the architecture
- paste the requirements
- paste the coding rules
- paste the brand voice
- paste the customer notes
- remind the model what was decided yesterday
You're doing the remembering: manual state management.
Why it fails
A transcript is not memory. Long conversations mix together:
- temporary context
- permanent decisions
- stale decisions
- entities
- preferences
- facts
- abandoned plans
The model has no clean way to tell which is which. That's how an agent recommends the file path you deleted three weeks ago: the wrong kind of context is available.
The remedy
Memory should be routed across separate indexes.
In Context Runtime, agent memory is split into three retrieval surfaces:
-
Recency
What just happened. -
Semantic memory
Durable decisions and project knowledge. -
Entity memory
People, systems, products, customers, files, concepts.
Different requests need different memory: a follow-up needs recency, a design question semantic, a customer question entity. Searching all three every time wastes tokens and increases drift.
What we measured
A learned routing policy improved reward by +2.9 compared with always searching every memory index. The win included cost and reduced irrelevant memory injection.
The point
Memory is retrieval selection from storage.
The storage question:
Where do we store history?
The hard question:
What deserves to be remembered for this request?
That is a planning problem.
Reference implementation: https://github.com/redevops-io/context-runtime
Benchmarks: https://redevops.io/benchmarks
Whitepaper v2: https://redevops.io/whitepaper-v2
redevops.io builds rAIlsApps (source-available (AGPL-3.0 with the Commons Clause)), self-hostable modules run by agents, on hardware you can actually afford. See it run.
