Skip to content
Platform · Self-Improving

The Platform Gets Sharper Every Day

A 3-tier self-learning pipeline that captures implicit corrections, clusters them semantically, promotes the recurring patterns, and auto-prunes the ones that stop working. Plus a unified context search across 13 scopes in a single call.

Four Stages, One Loop

The pipeline turns your corrections into durable behavioral rules — and discards them when they stop working.

STAGE 1

Implicit Capture

When you regenerate a slide, retry a failing tool, or correct an agent output, the platform notices. No thumbs-up / thumbs-down required — your actions are the signal.

STAGE 2

Semantic Clustering

Pgvector embeds every captured correction. When 3+ similar corrections (cosine ≥ 0.80) collect, a promotion manager distills them into one reusable behavioral rule.

STAGE 3

Theme Consolidation

At 30+ learnings, related rules consolidate into high-level themes. The agent prompt stays compact — broader applicability, fewer tokens, smarter behavior.

STAGE 4

Effectiveness Pruning

Learnings auto-archive when their success rate drops below 20% after 10+ uses. Knowledge gets sharper over time, not noisier — the opposite of how most prompt-engineering loops fail.

The Whole Learning Stack

Beyond the 4-stage pipeline — a skill curator, an LLM review queue, and effectiveness-aware retrieval.

Implicit Capture

Auto-detects dissatisfaction from user actions — slide regen, code re-runs, tool retries — no explicit feedback required.

Semantic Clustering

Pgvector embeds every correction. The promotion manager batches 3+ similar corrections (cosine ≥0.80) into one distillation pass.

Theme Consolidation

At 30+ learnings, related rules consolidate into high-level themes — fewer prompt tokens, broader applicability.

Effectiveness-Aware Pruning

Learnings that score below 20% success after 10+ uses auto-archive. Knowledge gets sharper, not noisier, over time.

Skill Curator

Pgvector-based dedup detection across the 19,000+ built-in skill library. Catches near-duplicates surface tokens miss.

LLM Review Pipeline

Automated keep / rehabilitate / merge / retire verdicts on every filesystem skill — operator-gated, never auto-applied.

Knowledge & Retrieval System

One coordinator, 13 context scopes. Hybrid search merges vector + keyword across memories, files, KDS, learnings, corrections, wiki, and project knowledge — deduped and ranked in a single call.

70/30 RRF

Hybrid Search

Reciprocal Rank Fusion combining vector similarity (70%) and BM25 keyword matching (30%) for optimal retrieval.

13 Scopes

13 Unified Context Scopes

One search hits memories, files, KDS, learnings, corrections, wiki, company docs, project notes / errors / summaries — deduped and ranked.

90-day half-life

Temporal Decay

Knowledge relevance decays with a 90-day half-life, ensuring recent information is prioritized in search results.

<7 day boost

Recency Boost

Documents less than 7 days old receive an additional boost, surfacing the most current knowledge first.

Code-Aware Chunking

Intelligent text chunking that respects code boundaries, function signatures, and document structure for cleaner embeddings.

Context-Aware Embeddings

Embedding generation that incorporates surrounding context and metadata for higher-precision semantic matches.

What This Means in Plain English

You correct the agent once. Slide regenerated? Tool retry? Code re-prompted? That's a signal. The agent doesn't need you to explain why — it embeds the moment of friction.

When the same pattern shows up three times, the platform distills it into a portable rule and injects it into the agent's context next time you (or anyone with similar work) start a related task. Compounded across thousands of users, the agent gets domain-aware without you ever opening a settings page.

When a rule stops helping, it auto-archives. The agent doesn't accumulate stale prompts — the opposite of how most agent platforms degrade over time.

And when you ask a question, the unified search hits memories, project notes, learnings, wiki pages, file content, code index, and 7 more scopes in a single call. One coordinator, parallel fan-out, deduped and ranked.

An Agent That Earns Its Keep

Self-improvement is a feature, not a roadmap promise. The platform shipped with this loop running.