Project releases
AgentDB 3.0.0-alpha.20 gives agents 41 memory tools over MCP
AgentDB is persistent memory for agents on top of the RuVector engine. The alpha.20 release ships the governed embedding identity work and a 41 tool MCP surface you can register with one command.
GitHub activity: · Published:
What AgentDB is
An agent that forgets everything between sessions repeats its mistakes. AgentDB is the store that stops that: patterns, skills, reflexions and causal edges, on the Rust RuVector engine with native Node bindings.
What changed
The alpha.20 release commit follows two changes worth knowing about: governed embedding identity with MetaHarness promotion, and a fix preventing lost updates in the sql.js path along with MCP stdout pollution.
Stdout pollution is an unglamorous bug with a large blast radius — anything printed on stdout corrupts the MCP protocol stream and the host sees a broken server rather than a bad log line.
Get started
Node.js. Three paths; the CLI needs no install:
npx agentdb init my-memory.rvf
npx agentdb add my-memory.rvf "vector memory that learns"
npx agentdb search my-memory.rvf "self-improving search" --top-k 5
Expected result: a local .rvf store and a ranked result for your query.
To make it callable from Claude Code or another MCP host:
claude mcp add agentdb -- npx agentdb@latest mcp start
That registers 41 tools across six families: pattern store and search, skill library, reflexion, causal edges and hierarchical recall.
Use it today
Practical case: a team memory for a codebase. Input is short notes about what worked; the workflow is store during review, search before starting; the output is a specific prior example instead of a generic suggestion.
Acceptance test: add five entries, then search with a phrase that shares no words with any of them. A useful hit means embeddings are working; no hit means you are still doing keyword search.
Push it further
Experimental commentary. Skill composition — chaining stored skills A to B to C with a bandit picking the composition — is the ambitious part of the API. The limitation is that bandits need enough signal to be better than a fixed order.
Falsifiable test: compare bandit chosen composition against a hard coded chain on the same fifty tasks. If the bandit is not ahead, you do not have enough signal yet.