Use cases
For agents that have to be right twice
The same failure everywhere: an agent that was right last month is confidently wrong today, and nothing in the transcript shows which. A bounded prompt and a recorded reason for every line are what turn one session into continuity.
Coding agents
Hold the build id, the deploy command, the naming convention, the thing that broke last month. Get them back in the one line that matters.
- Tends to hold
- Build ids, hosts, paths, conventions, deploy commands
- What tiering buys
- The window stays free for the diff, because a fact you did not name costs a line and not a body.
Support agents
Remember what this customer was told, what was actually true, and which of the two is now contradicted.
- Tends to hold
- What was said, what was true, ticket-level commitments
- What tiering buys
- A correction is pinned into the next prompt, so the agent does not repeat a promise that was withdrawn.
Research assistants
Keep a running set of findings and the open questions, with disagreements surfaced rather than averaged away.
- Tends to hold
- Findings, sources, open questions, superseded claims
- What tiering buys
- Two sources that disagree surface as a question instead of one of them quietly winning.
Internal assistants
Company facts with a source, an owner, and a date — and a running record of which topics the assistant is weak on.
- Tends to hold
- Runbooks, ownership, policy, tribal knowledge
- What tiering buys
- A domain the assistant keeps failing in gets a guardrail in every prompt until it earns its way back.
Long-running work
A service the agent can call for the hundredth session. Memory survives deploys, restarts, and model swaps.
- Tends to hold
- Everything learned, across sessions and agents
- What tiering buys
- It is a service, not a file in the repo — so it is queryable, revocable, and per-tenant.
Multi-tenant products
Memory as a credentialed storage layer: your own database, your own keys, revocable per integration.
- Tends to hold
- Per-organisation facts, isolated by key
- What tiering buys
- Every read is filtered by organisation in one auditable place, so the boundary is a single check.
If none of these fit
The requirement underneath all six
You need a store where a fact has a per-turn price, a disagreement is a record rather than a coin toss, and you can read every byte your agent has been told. If that is the shape of your problem, the use case is incidental.
Read on