Projects & memory

Memory & documents

Project memory is a searchable list of short entries that capture what a project knows and decided. Humans and agents write to it; agents read it before they act, so knowledge does not get lost between chat sessions.

Memory types

Type Use it for Example
Fact Something verified. Customer churn in Q2 was 3.1%.
Decision A choice and why it was made. Poland CPI is prioritized over Germany because of data quality.
Assumption Something believed but not verified. Average order value stays under €80.
Research Findings, sources, links. Eurostat publishes monthly HICP with a 2 week delay.
AI summary A summary written by an agent. Summary of the 12 interviews in the Feedback table.
Note Anything else. Ask finance for the 2025 budget file.

Each entry has a title, optional details, tags and a source (a user, an agent name or an import).

Adding memory

In the UI: open the project’s Memory tab, choose a type, fill in title, details and comma-separated tags, and click Add.

With an agent: ask “Remember that we chose weekly granularity for the revenue report.” The agent calls remember. It can also correct entries (update_memory) or remove outdated ones (forget_memory). Like every agent write, these follow the project’s autonomy level.

Searching

The search box on the Memory tab searches by meaning, not just by words: “why did we pick Poland?” finds the decision above even though it does not contain the word why.

  • Semantic mode is used when an embeddings service is configured (BASEROW_EMBEDDINGS_API_URL). Results are ranked by similarity.
  • Text mode is used otherwise; results match words in the title, content or tags. The page says which mode was used.
  • New entries are embedded in the background within seconds. Until then they are still found by text.

Filter by type with the dropdown next to the search box.

Agents use search_project_memory with the same behaviour. The API endpoint is GET /api/projects/<id>/memories/search/?q=....

Documents

Documents are longer texts attached to a project: meeting notes, briefs, specifications. They are included (as titles and excerpts) in the project briefing returned by get_project_context.

Good habits

  • Write one idea per entry and use a clear title; the title is what agents see first.
  • Record decisions with their reason. The reason is what future you and future agents need.
  • Mark guesses as assumptions and turn them into facts once verified.
  • Use tags such as pricing, q3, data-quality to filter quickly.

Re-indexing

After changing the embeddings service, rebuild the vectors with:

./baserow reindex_project_memories            # all projects
./baserow reindex_project_memories --project 12 --missing-only