reasoningbank-intelligence
Skill19 starsChanged 3 months agoMITView on GitHub
Adaptive learning for moflo agents via ReasoningBank: trajectory storage, verdict judgment, memory distillation, consolidation, and MMR retrieval. Use when building agents that should improve from experience across runs.
In this repository30 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- src/CLAUDE.mdCLAUDE.md
- communeSkill
- connector-builderSkill
- distillSkill
- divineSkill
- eldarSkill
- flflSkill
- flo-simplifySkill
- floSkill
- floSkill
- guidanceSkill
- healerSkill
- luminariumSkill
- meditateSkill
- memory-optimizationSkill
- memory-patternsSkill
- memory-teamSkill
- memory-worktreeSkill
- optimize-learningsSkill
- perf-auditSkill
- publishSkill
- quickenSkill
- reset-epicSkill
- spell-builderSkill
- spell-scheduleSkill
- test-gapsSkill
- vector-searchSkill
- verifySkill
- wardSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 120
Don't share
namespacebetween different agent roles: Keep reasoning-bank:reviewer, reasoning-bank:researcher, etc.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
ReasoningBank Intelligence
Trajectory-based learning pipeline for moflo-enabled agents. Records what an agent did, judges the outcome, distills successful runs into reusable patterns, and retrieves relevant prior experience on the next task.
Prerequisites
- moflo (the neural subsystem ships inline as part of the moflo package)
- Moflo's memory DB at
.swarm/memory.db(created on first run)
Quick Start
import { createInitializedReasoningBank } from 'moflo/dist/src/cli/neural/reasoning-bank.js';
const rb = await createInitializedReasoningBank({
namespace: 'reasoning-bank',
vectorDimension: 768,
retrievalK: 3,
mmrLambda: 0.7, // 0=pure relevance, 1=pure diversity
distillationThreshold: 0.6, // min verdict score to keep
dedupThreshold: 0.95,
});The Pipeline
Four stages. You typically call each once per task:
1. Record trajectory → storeTrajectory({ id, input, actions, outcome, reward, ... })
2. Judge → const verdict = await rb.judge(trajectory)
3. Distill → const memory = await rb.distill(trajectory) // if verdict good enough
4. Retrieve (next task) → const hits = await rb.retrieveByContent(query, k)1. Record
const trajectory = {
id: taskId,
input: userRequest,
actions: ['read_file', 'edit_file', 'run_tests'],
outcome: 'success' as const,
reward: 1.0, // 0..1
metadata: { toolCalls: 3, durationMs: 1800 },
timestamp: new Date(),
};
rb.storeTrajectory(trajectory);2. Judge
The built-in judge scores on outcome + reward + action-step quality. No external LLM call.
const verdict = await rb.judge(trajectory);
// { score: 0-1, outcome: 'success' | ..., reasoning: string }Swap in your own judge by extending ReasoningBank if you want LLM-in-the-loop scoring — this was designed as a rule-based baseline so the hot path stays cheap.
3. Distill
Compresses a trajectory into a reusable DistilledMemory (signature, approach, outcome-tagged). Skips if the verdict is below distillationThreshold.
const memory = await rb.distill(trajectory);
// null if verdict below threshold → not worth learning fromFor batch runs (nightly, offline replay):
const memories = await rb.distillBatch(trajectories);4. Retrieve
Query by text (embedding generated for you) or by pre-computed vector:
const hits = await rb.retrieveByContent(newTaskDescription, 5);
// [{ memory, relevanceScore, diversityScore, combinedScore }]MMR (maximal marginal relevance) prevents the top-K from collapsing to "five slight variations of the same thing" — tune mmrLambda to push more diversity.
Consolidation
Memory quality degrades over time without maintenance. Run consolidation periodically (once a week, once after N new entries, etc.):
const result = await rb.consolidate();
// { removedDuplicates, contradictionsDetected, prunedPatterns, mergedPatterns }Consolidation:
- Removes duplicates above
dedupThresholdsimilarity. - Detects contradictions (same signature, opposite outcomes) if
enableContradictionDetection. - Prunes entries older than
maxPatternAgeDays. - Merges semantically-adjacent patterns.
Persistence
ReasoningBank persists through MofloDbAdapter to .swarm/memory.db. Set enableMofloDb: false for ephemeral in-memory use (tests).
The namespace config isolates reasoning-bank entries from general memory. Default is reasoning-bank.
Anti-Patterns
- Don't record every step as a separate trajectory. A trajectory = one task. Steps are a field inside the trajectory.
- Don't skip the judge. Distilling every trajectory poisons the pool with failures — that's what
distillationThresholdguards against. - Don't run consolidation on the hot path. It's a sweep — do it out-of-band.
- Don't share
namespacebetween different agent roles. Keepreasoning-bank:reviewer,reasoning-bank:researcher, etc. separate; signatures overlap otherwise. - Don't raise
retrievalKto tame bad retrieval. TunemmrLambdaor the embedder instead — more K just dilutes the top results.
Integration with moflo's Hooks
Moflo's session/hook system already wires ReasoningBank into the /flo spell and the SubAgentStart hook. If you're building a custom agent that should participate, hook into post-task:
await mcp.hooks_post_task({
trajectoryId: taskId,
outcome: 'success',
reward: 1.0,
});That records, judges, and distills in one call.
Performance
- Retrieve (k=5, corpus of 10k): ~3–8ms.
- Distill: single call ≈ vector embed + a few similarity checks; ~10–50ms.
- Consolidate: O(n²) in the namespace — run offline for corpora > 10k.
See Also
memory-patternsskill — for non-trajectory memory (sessions, knowledge)memory-optimizationskill — HNSW tuning, quantizationsrc/cli/neural/reasoning-bank.ts— full APIsrc/cli/services/learning-service.ts— trajectory types
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

