Writing
Why agent memory is an infrastructure problem
Chat logs and vector dumps are not memory. Durability, schema, provenance, and evaluation are.
Memory becomes useful when it behaves like infrastructure: durable, typed, queryable, and honest about where each fact came from.
The mistake
Teams bolt “memory” onto agents by concatenating recent messages or embedding everything that happened. That produces recall-shaped behavior and none of the properties you need in an operational system: inspectability, provenance, and a defined failure when the record is missing.
What memory has to be
- An append-only evidence log (what happened).
- Typed projections (failures, interventions, decisions, observations).
- Retrieval that can filter on identity and time, not only similarity.
- A write path that labels the agent’s own outputs so they do not silently pollute the next retrieve.
Why this is infrastructure
If you cannot rebuild memory from the log, you do not have a system. You have a cache with vibes. If a retrieved item cannot point at a source, you cannot defend a decision. If evaluation only scores fluent answers, you will ship a storyteller.
- Asset dataState & telemetry
- EventsAppend-only evidence
- MemoryTyped projections
- RetrievalQuery + provenance
- ReasoningBounded context
- DecisionWritten back as memory
Open questions I still want to write through
- When is summarization an index rather than a source of truth?
- How should memory be scoped across assets, tenants, and time?
- What does a golden retrieval set look like for an industrial agent?