ReasoningBank learning with agentic-flow

ReasoningBank stores one episode per finished task (task text, reward, success, critique) in AgentDB and finds similar episodes by vector search. The loop is retrieve, act, judge, store.

Which tools exist depends on the MCP profile

The plugin's mcp_profile setting (default readonly) decides what the agentic-flow server registers:

Profile	Search and stats	Store an episode	Spawn an agent
`readonly`	yes	no	no
`retrieval`	yes	no	no
`learning`	yes	yes	yes (`agentic_flow_agent`)
`admin`	yes	yes	yes

Tools appear as mcp__plugin_agentic-flow_agentic-flow__<tool>. Claude Code asks before each call; the plugin pre-approves none.

1. Retrieve

Call agentdb_pattern_search with:

task: the new task in one sentence
k: 3 to 5
onlySuccesses: true for approaches to reuse; onlyFailures: true for mistakes to avoid

Read the critiques, not just the rewards.

2. Act

Do the task. Keep the retrieved critique in mind.

3. Judge

Score honestly: reward 0 to 1 and success true or false. A test suite that passes is evidence; "it looks right" is not.

4. Store (learning or admin profile)

Call agentdb_pattern_store with sessionId, task, reward, success, and a two or three sentence critique. Leave out secrets, keys and personal data.

Trends

agentdb_pattern_stats (by task) and agentdb_stats (database) report what is stored. Quote their numbers; do not extrapolate them.

From a terminal

npx agentic-flow reasoningbank status, list, consolidate and help work outside Claude Code too.