Raava / Reasoning Bank

04 · Case study

A memory agents learn from.

The Reasoning Bank is the memory our Foundry agents work from: what worked, what failed, and when.

InternalServed over MCP
The Reasoning Bank cycle: an agent session is traced, a judge turns the trace into a candidate lesson, a person accepts or rejects it, accepted lessons go into the bank, and agents recall them at the start of the next task. Agent sessionon Foundry Traceevery step kept Judgescores the attempt Candidatedo or avoid, and when A personaccepts or rejects The bankreviewed lessons recalled over MCP at task start and when a tool fails
Only lessons a person accepted are ever recalled.

From traces to lessons

Every agent session is traced. A pipeline inside Foundry reads those traces, rebuilds what the agent tried, judges how it went, and writes a candidate lesson: what to do or avoid, and when. A person accepts or rejects each one. Only accepted lessons are kept.

Recalled when it matters

Agents search the bank over MCP when a session starts and again when a tool call fails, so the next attempt begins from evidence instead of from zero. Search uses embeddings and a reranker. On real agent tasks the reranker found the right lesson 96% of the time, against 68% without it.

Small on purpose

It is internal. Of the first 26 candidate lessons, reviewers kept 15.

Building something agents should run?

Raava Solutions takes on Palantir Foundry and AI agent work for clients. Tell us about the operation.

contact@raavasolutions.com