How Volary works
Volary is a memory system for your AI agents.
It lets them automatically learn from past experience and recall that knowledge in future sessions, so they pick up where they left off.
The memory system
Episodes
Volary collects & categorises agent transcripts and conversations into episodes, summarised pieces of information about their interactions. These present learning opportunities:
- What went wrong – Times when something was unsuccessful, the agent performed badly, or took too long.
This can be direct user feedback, or learned implicitly from the transcript, such as tool failures. - Learned experiences – Cases where the agent discovered something useful, or changed its approach because of what it learned.
Reflections
Volary seeks through these episodes and distils reflections from them, which are learnings that the agent can reuse in future. These come in different flavours:
- Factual– Concrete facts that the agent had to learn during the conversation, which it will be able to recall next time, e.g.
- “Unit tests are co-located with source files in this repo”
- “Users prefer it when I return them direct links to documentation rather than waiting for them to ask”
- Procedural– Tactical advice to advise the agent how to make more strategic choices in future, e.g.
- “When searching for specific vulnerabilities, I should use a language-specific search engine when available (such as https://pkg.go.dev/vuln) rather than Google”
Integration
You can configure Volary as a backend for your agent, simply by changing its API URL and adding the Volary MCP tools. No code changes are needed, and Volary will start reflecting on your conversations immediately. Volary supports sending requests to all major AI backends.
If you're building an agent yourself, you can directly integrate Volary as tools for your agent.
An SDK is coming soon to make this even easier!
Get started
What are you working on?