The short answer
A correct answer is only useful evidence when you can inspect its source.
See what the local retrieval test returned, how to repeat it and what it did not test.Your AI recalls a delivery date from an earlier task. Ask which source supports it before acting. This test checks whether the date survives the session, whether you can inspect the evidence and how retrieval responds when a requested fact is missing.
You can adapt this small test to your own agent workflow. We tested storage and retrieval against a Gigabrain release archive with fictional project data, using separate local processes in lexical recovery mode. It does not establish how a hosted agent will reason, how the default model-backed mode will perform or how every future query will behave.
Start with one fact you can check
Use a dedicated test folder and a note whose answer you already know. Our fictional source said:
Aster supplies Project North. The current delivery date is 12 October.
The fictional project avoids exposing private work. The expected date gives you a direct way to check both the answer and its citation. Do not start by importing a whole company archive or asking the system to judge its own accuracy.
Test persistence in a separate process
- Extract the release into isolated temporary storage.
- Add the fictional source to the kit’s raw source folder.
- Run ingestion and build the local index.
- Query for
Project North deliveryusing explicit lexical mode and save the JSON response. - Run the query again from another process and inspect the returned quote, source, line and citation handle.
Running the query in a new process checks that the information was saved to files and can be retrieved again. You also need to test a fresh agent session: does your integration call the right tools, and does the agent use the results correctly?
Current result status
The published component result is awaiting release freshness verification. No passing result is asserted for the current archive. Use the method above and inspect the evidence page for its current status.
Include a question the notes cannot answer
We also asked about a warranty term that was absent from the fictional notes. The tool returned no finding. A fluent answer can otherwise hide the missing source. When you repeat this in an agent conversation, ask the agent to identify the gap and tell you what additional source it needs.
A refusal on one absent query is not a general accuracy score. Repeat with relevant missing facts, ambiguous entity names and conflicting records. Keep the expected result and observed result separate so you can inspect failures instead of counting every plausible sentence as a success.
Four checks to run before relying on memory
| Check | What to inspect |
|---|---|
| Persistence | Can a later process or session recover the required information? |
| Provenance | Does the cited file and line actually support the answer? |
| Correction | After a dated amendment, does the workflow distinguish the current fact from its history? |
| Isolation | Can an unrelated project answer only from its permitted sources? |
We tested persistence, a source citation and one missing-evidence case. The correction and project-isolation checks above are suggested next steps. This test did not cover them. For integration-specific evidence, read the tested agents and versions.
Is this the kind of memory system you need?
Gigabrain combines source ingestion, semantic and lexical retrieval, and source-linked knowledge graphs for your AI agent. You get the indexing, retrieval and verification tools in a system you run locally. You choose the sources, agent and permissions. If you need a provider to host and operate memory for multiple customers, compare managed services before choosing a system you operate locally.
Start with the persistent-memory guide if you are defining your workflow. If that suits your work, check your agent’s setup, read how the system works and use the plan comparison below.
Method note: this article was prepared with AI assistance from the recorded synthetic test and the release evidence. The downloadable result is the source for the observations above. Gigabrain Labs publishes this field note about its own product.
Add a connected-source correction check
The lexical component result above checks a later process and one missing-evidence query. A separate current demonstration checks a connected schedule through the local memory service.
We reproduced the connected-source run against both 1.1.0 editions. After its disposable schedule changed from 18 October to 25 October, the old citation failed verification. Updated support restored the path to the new passage.
Inspect the correction record. These are local component results. Neither demonstration certifies a fresh conversation with every supported agent.
Record the fresh-session boundary explicitly
- Start a new agent conversation.
- Avoid pasting the expected answer into the prompt.
- Ask the agent to retrieve the project fact.
- Save the actual tool invocation.
- Compare the answer with the returned passage.
- Record any unsupported inference.
Use the fresh-session record for your integration. Check the dated compatibility evidence separately. A service result does not establish that an agent invoked the service.
Method note: Gigabrain Labs prepared this material with AI assistance. Fictional teaching examples and executed results are labelled separately.
